0 Bewertungen0% fanden dieses Dokument nützlich (0 Abstimmungen)

29 Ansichten38 SeitenApr 08, 2013

© Attribution Non-Commercial (BY-NC)

PDF, TXT oder online auf Scribd lesen

Attribution Non-Commercial (BY-NC)

Als PDF, TXT **herunterladen** oder online auf Scribd lesen

0 Bewertungen0% fanden dieses Dokument nützlich (0 Abstimmungen)

29 Ansichten38 SeitenAttribution Non-Commercial (BY-NC)

Als PDF, TXT **herunterladen** oder online auf Scribd lesen

Sie sind auf Seite 1von 38

CHRISTINE A. BOTOSAN, University of Utah MARLENE A. PLUMLEE, University of Utah HE WEN, University of Utah 1. Introduction Existing literature employs two general approaches to assess the validity of alternative proxies for rm-specic cost of equity capital or expected return (hereafter Et)1(rt)). The rst approach involves examining the association between the proxy for Et)1(rt) and future realized returns. The second approach focuses on the association between the Et)1(rt) proxy and contemporaneous risk characteristics of rms. The results of these two streams of literature are mixed. Easton and Monahan (2005) (hereafter EM) and Guay, Kothari, and Shu (2005) (hereafter GKS) focus on the association between alternative proxies for Et)1(rt) and future realized returns and conclude that none of the proxies they examine provide valid estimates of the construct of interest. In contrast, Botosan and Plumlee (2005) (hereafter BP) conclude that two common proxies for Et)1(rt) rDIV (Botosan and Plumlee 2002) and rPEG (Easton 2004) are valid, based on their nding that both are associated with rm-specic risk characteristics in a theoretically predictable and stable manner. Furthermore, Pastor, Sinha, and Swaminathan (2008) document a positive association between market-level implied cost of capital and risk as measured by the volatility of market returns, consistent with the estimates capturing time-varying Et)1(rt). In this paper, our goal is to reconcile the conict between these two streams of literature and provide additional evidence pertaining to the construct validity of the proxies employed in extant research. Contrary to the results documented in EM and GKS, we document a positive association between ten of the twelve Et)1(rt) proxies included in our study and future realized returns after controlling for new information.1 We reconcile our ndings to those in EM and GKS by demonstrating that the prior results are due to empirical misspecication. Finally, we show that two of the proxies, rDIV and rPEG, demonstrate not only the expected relation with future realized returns, but also with rm-specic risk. We also address several other issues regarding the use of implied cost of capital estimates including: (1) analysts forecast bias, (2) the efcacy of realized returns for Et)1(rt) before and after controlling for news, (3) the effectiveness of averaging several Et)1(rt) proxies, and (4) the substitution of realized values for analysts forecasts of cash ows or earnings. Our evidence suggests that deviations between analysts expectations and those of the market lead to potentially less powerful proxies but do not generate biased or

* Accepted by Steven Salterio. We gratefully acknowledge the nancial support of the David Eccles School of Business. We also wish to thank Kin Lo, K. Ramesh, Matt Magilke, and the workshop participants at the London School of Business, Michigan State University, Rotterdam School of Management Erasmus University, the University of North Carolina, the University of British Columbia, Arizona State University, Georgetown University, and University of Akron for helpful comments on previous drafts of the paper. 1. Of the twelve estimates we examine, nine are implied cost of capital proxies, two are popular averages of subsets of these proxies, and the nal estimate is derived from the Fama-French four factor model.

Contemporary Accounting Research Vol. 28 No. 4 (Winter 2011) pp. 10851122 CAAA doi:10.1111/j.1911-3846.2011.01096.x

1086

inconsistent results. Furthermore, we nd that realized returns do not proxy for Et)1(rt) even after controlling for news, and that averaging several proxies does not yield an enhanced metric. Finally, substituting realized values for analysts forecasts of cash ows yields systematically biased estimates, which might yield biased and inconsistent results when such estimates are employed in empirical research. Given the current state of the literature, the validity of the various cost of capital estimates is unclear and it is not uncommon for similar studies to document dissimilar results because they employ different cost of capital estimates. For example, Ogneva, Subramanyam, and Raghunandan (2007) (hereafter OSR) conclude that rms with internal control weaknesses do not bear a higher cost of equity capital, while Ashbaugh-Skaife, Collins, Kinney, and LaFond (2009) (hereafter ACKL) conclude the opposite. These contradictory results are wholly attributable to differences in the authors choices of Et)1(rt) proxies.2 Thus, additional evidence regarding the validity of alternative Et)1(rt) proxies is needed to help guide researchers proxy selection. Based on our evidence, we recommend that researchers requiring a valid proxy for Et)1(rt) employ either rDIV or rPEG. We caution against the use of realized returns or the other implied cost of capital estimates we examine to proxy for Et)1(rt). Finally, we suggest that researchers introducing new proxies for Et)1(rt) to the literature subject their proposed measures to both of the construct validity tests employed in the current study, and provide support for how the measure enhances the existing technology for estimating Et)1(rt). Our ndings should be of interest to researchers requiring a valid proxy for Et)1(rt). The need for such a proxy is far-reaching. It extends from accounting research that examines the impact of nancial reporting and disclosure on required returns, to nance research that investigates market anomalies, asset pricing theory, and capital budgeting. These areas of research produce ndings of interest to standard setters, regulators, investors, preparers, and auditors. Accordingly, the validity of the Et)1(rt) proxies employed in research is an important issue with pervasive implications. We organize the remainder of our paper as follows. Section 2 reviews the related literature and sets forth the development of our hypotheses. Section 3 presents our empirical models and proxies. Section 4 delineates our sample selection procedure and provides descriptive statistics for our sample. We discuss the results of our construct validity assessment in section 5. In section 6 we provide evidence pertaining to several other issues related to implied cost of capital estimates. Section 7 reconciles the results of our realized return analysis with those documented in prior literature. Finally, section 8 summarizes our conclusions. 2. Literature review and hypotheses development

Construct validity is woven into the theoretical fabric of the social sciences, and is thus central to the measurement of abstract theoretical concepts.. .. Fundamentally, construct validation is concerned with the extent to which a particular measure relates to other measures consistent with theoretically derived hypotheses concerning the concepts (or constructs) being measured.

2.

OSR replicate ACKLs ndings using ACKLs proxy for cost of equity capital, but discount the results, citing conicting evidence regarding the relative validity of alternative implied cost of capital measures. Our results provide little support for the Et)1(rt) proxy employed in OSR, but strong support for the proxy employed in ACKL.

1087

The preceding quote describes the classic, well-accepted method routinely employed to examine construct validity.3 Consistent with standard practice, we rely on the preceding quote to guide our approach to examining construct validity. Specically, we examine the relationship among alternative proxies for Et)1(rt) and future realized returns, which are a function of Et)1(rt). In addition, we examine the relationship among alternative proxies for Et)1(rt) and contemporaneous risk characteristics nance theory suggests are predictably associated with Et)1(rt). The following paragraphs detail the theories underlying our examination. Realized returns Realized return at time t (rREALt) is modeled as the expected return at time t conditional on information available at t ) 1 (Et)1(rt)) plus the unexpected (or abnormal) return due to new information (URt): rREAL;t Et1 rt URt 1:

Relying on prior research (e.g., Campbell 1991;Vuolteenaho 2002) further decomposes the unexpected return to new information (URt) into two components the unexpected return due to cash ow news (Ncf,t), and the unexpected return due to expected return news (Nr,t). This gives rise to equation 2: rREAL;t Et1 rt Ncf ;t Nr ;t where: rREAL,t = realized return from t)1 to t; Et)1(rt) = expected return at t, conditional on information at t)1; Ncf,t = return due to cash ow news from t)1 to t; and Nr,t = return due to expected return news from t)1 to t. Traditionally, research that employs realized returns to proxy for expected returns relies on the assumption that URt is mean zero, and that in-sample averaging of realized returns across rms or time purges URt to produce a valid proxy for Et)1(rt). Some more recent research goes further by using rm-specic (i.e., not averaged) realized returns to proxy for Et)1(rt) (e.g., Easley, Hvidkjaer, and OHara 2002; McInnis 2010). Nevertheless, a growing body of research questions the validity of realized returns as a proxy for Et)1(rt). Elton (1999) argues that averaging does not eliminate URt because unexpected returns tend to be large and correlated across rms and time. Vuolteenaho (2002) demonstrates that the unexpected component is the dominant factor driving rmlevel stock returns, and that cash ow news is largely rm-specic, whereas expected return news is linked to systematic macroeconomic factors. Consistent with the latter nding, Campbell and Ammer (1993) nd that expected return news drives aggregate stock returns. These ndings suggest that rm-level and portfolio-level realized returns could be poor proxies for Et)1(rt). At the rm level the URt due to rm-specic cash ow news, as well as the URt due to systematic expected return news, contaminate the realized return proxy. At the portfolio level sufcient averaging might mitigate the URt due to rm-specic cash ow news, but it is less likely to mitigate the URt due to systematic, macroeconomic expected return news. Moreover, if cash ow news is correlated across rms and or time, averaging, even over large numbers of rms or long periods, might not purge either

3. Carmines and Zellner (CZ) 1979: 23; bolding added. Because theory rarely models abstract theoretical concepts completely, construct validation does not require the identication of an exhaustive list of the observable measures believed to be associated with the underlying unobservable construct. On the contrary, a theoretical basis for hypothesizing a directional association between the empirical proxy of interest and some set of measures associated with the underlying unobservable construct is paramount to credible construct validation.

2;

1088

component of the URt. Attempts to mitigate the problem by averaging over increasingly larger samples or longer periods invoke unpalatable stationarity assumptions (Chan and Lakonishok 1993). (2) and the research discussed in the preceding paragraph suggest that an examination of the association between rREAL,t and Et)1(rt) is vulnerable to omitted variables if Ncf,t and Nr,t are ignored. If Et)1(rt) is correlated with Ncf,t or Nr,t the resulting correlated omitted variable bias could result in biased and inconsistent results. Even if Et)1(rt) is not correlated with Ncf,t or Nr,t, however, omitting the latter two variables reduces the power of the analysis. In this case, no statistically signicant correlation between rREAL,t and Et)1(rt) might be observed even if one exists. (2) suggests that if an Et)1(rt) estimate is a valid proxy, we should observe a positive correlation between the proxy and rREAL,t after controlling for Ncf,t and Nr,t. This gives rise to our rst hypothesis.4 HYPOTHESIS 1. After controlling for Ncf,t and Nr,t, a positive correlation between a proxy for Et)1(rt) and rREAL,t provides support for the validity of that Et)1(rt) proxy. Risk characteristics Rosss 1976 Arbitrage Pricing Theorem (APT) models the expected return for a given period as a function of the risk free rate (rf) plus the risk premiums arising from K risk factors: XK Et1 rt rf ;t1 k E r rf ;t1 : 3 k1 k t 1 k Rosss APT does not identify the risk factors, although existing research suggests several candidates. The capital asset pricing model (CAPM) suggests that Et)1(rt) is increasing in market beta (Lintner 1965; Mossin 1966; Sharpe 1964). Modigliani and Miller (1958) support a positive association between leverage and Et)1(rt). Berk (1995) argues that market value of equity is systematically decreasing in priced risk such that Et)1(rt) is inversely related to the market value of equity and positively related to the book-to-price ratio. Finally, Beaver, Kettler, and Scholes (1970) assert that abnormal earnings arising from growth opportunities are inherently more risky, leading to a positive association between Et)1(rt) and growth. Equation 3 above and this body of research gives rise to our second hypothesis.5 HYPOTHESIS 2. A positive correlation between a proxy for Et)1(rt) and the risk free rate, market beta, leverage, book-to-price ratio and growth, and a negative correlation with market value of equity provides support for the validity of that Et)1(rt) proxy. Related empirical research Prior empirical research examines the association between alternative proxies for Et)1(rt), realized returns, and rm-specic risk characteristics. Nonetheless, as noted earlier, the

4. Our paper seeks to provide guidance to researchers seeking a valid proxy for Et)1(rt) by examining the extent to which implied cost of capital estimates proxy for expected returns (i.e., Et)1(rt)). Lee, So, and Wang (2010) provides guidance to those seeking to predict future returns by investigating the ability of certain implied cost of capital proxies to predict future realized returns (i.e., rREAL,t). These roles for implied cost of capital estimates are quite different. In the presence of cash ow news and expected return news, it is quite possible that a particular proxy might be a poor (good) proxy for Et)1(rt), but a good (poor) predictor of rREAL,t. Readers interested in a more in-depth discussion of these risk characteristics are referred to Botosan and Plumlee 2005.

5.

1089

ndings from this research are mixed. GKS regress realized returns on ve alternative proxies for Et)1(rt).6 The authors do not document the expected positive association between realized returns and their proxies, but their analysis does not control for Ncf,t or Nr,t. EM control for Ncf,t and Nr,t in their examination of the association between realized returns and seven proxies for Et)1(rt). Even so, EM nd that none of the proxies are positively associated with realized returns. GKS and EM conclude that their results do not provide support for the construct validity of the proxies they examine. In contrast, BP examine the association among ve proxies for Et)1(rt) and rm-specic risk characteristics. They conclude that rDIV and rPEG, are valid proxies for Et)1(rt) as both are correlated with rm risk characteristics in a theoretically predictable manner. While GKS, EM, and BP examine different sets of Et)1(rt) proxies, rPEG is examined by all three. Accordingly, the ndings of GKS and EM versus those of BP present an explicit conict in the evidence regarding this metrics construct validity. The conict with respect to rDIV is implied as opposed to explicit since neither GKS nor EM include rDIV in their analyses. In addition, since 2005 there has been an explosion in Et)1(rt) proxies employed in the literature, without a rigorous construct validity assessment of the same. Accordingly, the primary objectives of the construct validity portion of this study are threefold: rst, to investigate the source of the disparate results noted above; second, to augment BPs risk-based construct validity analysis of rDIV with a realized return analysis and extend their ndings over an additional period of time; and third, to use both the realized return and risk-based approaches to examine the construct validity of the more recent additions to the set of alternative proxies for Et)1(rt). 3. Empirical models and proxies Empirical method for realized return analysis Realized return model Our empirical specication of the return decomposition model (2) is given below. rREALit a0 b1 ERit 1 b2 CFN 1it b3 CFN TVit b4 EWER Nit b5 FSER Nit eit where: rREALit ERit)1 CFN_1it 4;

= 12-month buy and hold return, beginning the month after estimation of ER; = expected return proxy at t conditional on information at time t)1; = news about period t to t+1 cash ows received during the 12-month realized return period; CFN_TVit = news incorporated in target prices during the 12-month realized return period; EWER_Nit= economy-wide expected return news during the 12-month realized return period; and FSER_Nit = rm-specic expected return news during the 12-month realized return period. Recall that equation 2 models realized returns (rREAL,t) from t)1 to t as a function of expected returns at t conditional on information at time t)1 (Et)1(rt)), as well as cash ow news (Ncf,t) and expected return news (Nr,t) received from t)1 to t. In our specication ERit)1 is one of a number of alternative expected return estimates intended to proxy for Et)1(rt).7 Hypothesis 1 states that, after controlling for Ncf,t and Nr,t, a positive correlation between a given proxy for Et)1(rt) and rREAL,t (i.e., a positive 1 coefcient) provides support for the validity of that Et)1(rt) proxy.

6. 7. The Appendix summarizes the Et)1(rt) proxies examined in GKS, EM, BP, and the current study. We enumerate the expected return estimates included in our study in the Appendix and in section 3.

1090

In theory 1 should equal 1. Such a test is not only a test of the extent to which the proxy captures cross-sectional variation in Et)1(rt), but also the extent to which it captures the magnitude of Et)1(rt). Most empirical research employing Et)1(rt) proxies is concerned with cross-sectional variation in Et)1(rt). Accordingly, an Et)1(rt) proxy that captures cross-sectional variation in Et)1(rt) (i.e., 1 > 0) might be valid for use in empirical research even if the magnitude of the proxy is biased (1 1). Under such circumstances, a test of 1 = 1 is an unnecessarily rigorous test of construct validity. For this reason, in testing our rst hypothesis we do not require 1 to meet the more stringent test of equivalence to the theoretical value of one, but we report the results of this test. Our empirical specication includes two proxies for Ncf,t. CFN_1t captures cash ow news related to near-term cash ows, and is measured as the earnings surprise during the realized return period. Our second proxy, CFN_TVt, is the revision in analysts forecasts of target prices during the realized return period. We include this variable to capture cash ow news related to long-horizon cash ows. Since realized returns are increasing in cash ow news (see (2)), we expect 2 and 3 to be signicantly positive. Our model also incorporates two proxies for Nr,t. Since expected returns are a function of the risk free rate, we include the change in the risk free rate between t)1 and t to proxy for economy-wide expected return news linked to macroeconomic factors (EWER_Nt). Since expected returns are also a function of the amount of risk a particular rms stock presents, we include the change in rm-specic market beta between t)1 and t to proxy for expected return news linked to changes in the amount of risk associated with the rm (FSER_Nt). As shown in (2), expected return news is negatively associated with realized returns. Accordingly, we expect both 4 and 5 to be signicantly negative. Target prices, which we employ in the computation of CFN_TVt, reect the present value of innite horizon cash ows beyond the forecast horizon. As noted above, our primary objective for including CFN_TVt in our empirical model is to capture long-horizon cash ow news. Nevertheless, analysts revisions of target prices reect changing beliefs about discount rates as well as future cash ows.8 Since target prices are increasing in future cash ows but decreasing in the discount rate, we expect the association between CFN_TVt and realized returns to be positive regardless of whether CFN_TVt captures cash ow news, expected return news, or both. As both types of news are important to the theoretical model, CFN_TVt is a particularly important control variable. The following paragraphs provide details of the measurement of all variables included in (4) except for the Et)1(rt) proxies. The Et)1(rt) proxies are outlined following our discussion of the empirical model we employ in our risk analysis. Realized returns. We calculate realized returns using CRSP data as the buy and hold realized return computed over the 12 months beginning the month after we estimate expected returns. Cash ow news. To calculate our cash ow news proxies we rely on Value Line analysts forecasts of annual earnings per share for the current year (Year t) and target prices at the end of Year t+4. All forecasts are made in the third calendar quarter of the year. Our proxy for the current year cash ow news (CFN_1) is the difference between the reported annual earnings per share for year t announced during the 12-month period we estimate realized returns, less analysts forecasts of those annual earnings issued the day we estimate Et)1(rt). Thus, CFN_1 captures earnings surprise similar to that employed in numerous other studies. We compute CFN_TV as the difference between the midpoint of the forecasted target price range made 12 months after we estimate our expected return proxies and the midpoint of the forecasted target price range made at the date we estimate

8. Consistent with this, Lambert (2009) highlights that target prices are free to reect whatever assumption regarding future discount rates are deemed appropriate (261, note 1).

1091

our Et)1(rt) proxies. In our analyses, we scale our cash ow news variables by stock price on the day we estimate expected returns. We obtain actual earnings per share from COMPUSTAT and forecast and stock price data, which is stock price as of the publication date or within three days after publication, from Value Line. Panel A of Table 1 provides a description of the cash ow news proxies employed in our analysis. Expected return news. We compute our proxy for economy-wide expected return news (EWER_N) as the change in the risk free rate (rf) over the 12-month realized return period. Specically, EWER_N is the rf during the last month of the realized return estimation period less rf twelve months earlier.9 We measure rf as the ve-year treasury constant maturity rate obtained from the U.S. Federal Reserve at http: www.federalreserve.gov. Our proxy for rm-specic expected return news (FSER_N) is calculated as the change in market beta (MBETA) over the 12-month realized return period.10 We use CRSP data to estimate MBETA via the market model with a minimum of 20 out of 60 monthly returns and a market index return equal to the value weighted NYSE AMEX return. Each beta estimation period ends in June. Panel B of Table 1 provides a description of the expected return news proxies employed in our analysis. Empirical method for risk analysis Expected return model (5) provides our empirical specication of the expected return model given by equation 3: ERit 1 v0 c1 rft1 c2 UBETAit 1 c3 DMit1 c4 LMKVLit 1 c5 LBPit 1 c6 EXGRWit1 git 1 where: ERit)1 = expected return proxy; rft)1 = risk-free rate of interest; UBETAit)1 = unlevered CAPM beta; DMit)1 = leverage; LMKVLit)1 = log of the market value of common equity; LBPit)1 = log of book-to-market ratio; EXGRWit)1 = expected growth in earnings per share. Hypothesis 2 states that a theoretically predictable relation between a given proxy for Et)1(rt) and the risk-free rate of interest, market beta, leverage, market value of equity, book-to-price, and earnings growth provides support for the validity of that proxy. Specically, we expect ERit)1 to be increasing in the risk-free rate, market beta, leverage, bookto-price, and growth and decreasing in the market value of equity. Accordingly, nding c1, c2, c3, c5, and c6 greater than zero and c4 less than zero provides support for a given

9. Monahan and Easton (2010) question the use of this proxy, stating that the risk-free rate has nothing to do with risk and claiming that a change in rf is a cross-sectional constant. From (3), rf is an economywide parameter that bears directly on Et)1(rt). Thus, a change in rf gives rise to a change in Et)1(rt), such that a change in rf constitutes economy-wide expected return news, our construct of interest. In addition, rf is not a cross-sectional constant because, as stated in Table 1, we measure the change in rf using the ve-year treasury constant maturity rate as of the month the Et)1(rt) proxy is estimated. Monahan and Easton (2010) question the use of changes in MBETA to capture rm-specic expected return news. We employ the change in MBETA to proxy for rm-level changes in the amount of risk. In so doing, we do not presume that market risk is the only relevant priced risk, but we do assume that changes in MBETA are correlated with changes in the overall level of risk an investment presents to the market. The power of this proxy is reduced if this assumption is violated, but the potential detrimental effect of this is mitigated by the inclusion in our model of the change in expected terminal value, which also captures rm-specic expected return news.

5;

10.

1092

TABLE 1 Variable descriptions: cash ow news, expected return news, and rm-specic risk factors Panel A: Cash ow news proxies Variable CFN_1 Descriptiona = Current period cash ows news = (A_epsit eps1it) |Priceit|. Calculated as COMPUSTAT actual earnings per share for Year t+1 less the related Value Line forecast of Year t+1 earnings, made in the third quarter of Year t, scaled by the absolute value of the Value Line reported stock price at the time r is estimated (in Year t). = Terminal value cash ow news = (TPit+1 TPit) |Priceit|. Calculated as the Value Line forecast of target price made in Year t+1 less the Value Line forecast of target price made in Year t, scaled by the absolute value of the price at the time r is estimated (in Year t).

CFN_TV

Panel B: Expected return news proxies EWER_N = Economy-wide expected return news = Change in discount rate = rft+1 rft. Calculated as the ve-year treasury constant maturity as of the month Et)1(rt) in Year t+1 less the ve-year treasury constant maturity at the time Et)1(rt) is estimated (in Year t). rf is drawn from the U.S. Federal Reserve at http: www.federalreserve.gov. = Firm-specic expected return news = Change in market beta = mbetat+1 mbetat.

FSER_N

Panel C: Risk proxies MBETA = CAPM beta estimated via the market model using the value weighted NYSE AMEX market index return. Require a minimum of 20 monthly returns over the 60 months prior to June of the year Et)1(rt) is estimated. = unlevered CAPM beta = MBETA (1 + Debt Equity) where debt is long-term liabilities (COMPUSTAT items dltt) and equity is total common stockholders equity (COMPUSTAT item ceq) as of the end of the scal year prior to the date Et)1(rt) is estimated. = long-term liabilities (COMPUSTAT items dltt) at the end of the scal year prior to the date Et)1(rt) is estimated scaled by MKVL. = market value of equity on December 31st of the Year t)1 (prior to the date Et)1(rt)) is estimated. If the market value of equity from the Center for Research on Security Prices (CRSP) is not available use the one from COMPUSTAT as of the end of the scal year prior to the date Et)1(rt) is estimated, stated in millions of dollars. LMKVL is the natural log of MKVL. = book value of equity at the end of the most recent quarter prior to the date Et)1(rt) is estimated scaled by MKVL. LBP is the natural log of BP. = earnings growth computed by dividing the difference in forecasted earnings ve periods in the future less forecasted earnings four periods in the future by the absolute value of forecasted earnings four periods in the future.

UBETA

DM MKVL

BP EXGRW

Notes:

a

Throughout the description of the variables, Year t refers to the year that the expected return estimates (Et)1(rt)) are made.

1093

proxy for expected return. In theory c1 should equal 1, although a test of c1 = 1 is a particularly rigorous construct validity test. Nevertheless, we also report the results of a test of c1 = 1. We do not infer construct validity from the magnitude of the R2 of the cost of equity capital models because assumptions regarding the terminal value imposed by the researcher in the derivation of some implied cost of capital estimates can lead to induced spurious correlation between the proxy and certain risk characteristics, yielding a high R2 by construction. For example, if an assumption related to long-range growth in earnings is used to derive the terminal value (as in the PEG model), it is not surprising that growth, a rm-specic risk factor, explains a signicant portion of the variation in that expected return proxy. Thus, it is particularly important to assess the association between the expected return estimates and various rm-specic risk factors after controlling for risk factors that are potential candidates for spurious correlation (primarily growth). Finally, Berk (1995) argues that book-to-price and market value of equity play similar roles in an incomplete model of expected returns, in that both variables capture otherwise unmeasured risk. Accordingly, it is unclear whether both variables should achieve signicance in the empirical model. The catch-all nature of these variables, however, mitigates the concern that the model is susceptible to omitted risk factors. Below is a detailed discussion of our measurement procedures for the risk proxies employed in our analysis. Panel C of Table 1 summarizes their descriptions. Unlevered beta. We include unlevered beta (UBETA) to capture the theoretical relation between Et)1(rt) and CAPM beta. Including the traditional levered beta (MBETA) in the model captures leverage risk as well as market risk (e.g., Hamada 1972; Chung 1989), which complicates the interpretation of the coefcients on both leverage and beta. Employing unlevered beta circumvents this issue. We unlever MBETA using the procedure described in standard nance textbooks (e.g., Kester, Fruhan, Piper, and Ruback 1997).11 Briey,

Debti MBETAi UBETAi Equity UBETA;

i

which yields

MBETA is estimated as described earlier. We compute debt-to-equity by dividing long-term debt by stockholders equity using COMPUSTAT data measured at the end of the scal year prior to the time we estimate the Et)1(rt) proxies. Leverage. Modigliani and Miller (1958) suggest that as debt in a rms capital structure increases, risk increases. As discussed above, estimating the model with UBETA allows us to predict unequivocally a positive coefcient on DM (c3). We measure DM as the ratio of long-term debt (described above) to the market value of common equity measured on December 31 prior to the year we estimate expected return. We collect market value of common equity from CRSP, or from COMPUSTAT if the data are unavailable on CRSP. Market value of equity. Our proxy for the market value of equity is the natural log of the market value of common equity (LMKVL) and is collected as described above. Consistent with prior research, we log transform the data to mitigate skewness. Book-to-price. We calculate book-to-price (BP) as the book value per share at the quarter end closest to a date on or before June 30th of the Value Line publication year

11. The formula we employ to unlever beta assumes no certainty, and consequently no benet arising from the tax deduction of interest payments. The Hamada 1972 formula is an alternative approach that incorporates the value of the tax shield on interest by including an adjustment for tax (1)t) in the denominator. In the context of our analysis, there is no reason to expect the tax adjustment to play a role. Accordingly, we follow the approach typically used by investment banks and practitioners, and ignore the tax shield.

1094

scaled by price per share. Price per share is the stock price on the Value Line publication date or closest date thereafter within three days of publication. We log transform these data to mitigate skewness (LBP). Expected growth. We estimate expected earnings growth (EXGRW) using the forecasted growth in earnings ve years hence. We calculate EXGRW for each rm-year as the forecasted earnings for Year t+5 less forecasted earnings for Year t+4, scaled by the absolute value of the Year t+4 forecast. Before moving on to the measurement of the Et)1(rt) proxies, we emphasize the importance of timing in our analyses. As noted in (2), realized returns (rREAL) are a function of Et)1(rt) determined prior to the period over which realized returns and news are measured. Thus, in our test of Hypothesis 1 we measure realized returns, cash ow news, and expected return news contemporaneously, but our estimates of Et)1(rt) are measured at the end of the prior year. As noted in (3), however, rm-specic risks existing at t)1 are part of the t)1 information set investors use to determine Et)1(rt). Thus, in our test of Hypothesis 2, we employ risk proxies estimated contemporaneously with our Et)1(rt) proxies. Expected return proxies We analyze twelve proxies for Et)1(rt); all drawn from extant research. Nine are implied cost of capital proxies included in at least one of the three studies (GKS; EM; BP) that contribute to the debate regarding the validity of expected return estimates: rCT, rDIV, rGLS, rGOR, rOJN, rMPEG, rPEGST, rGM and rPEG (see the Appendix). Due to the popularity of this measure, we also include estimates derived from the Fama-French four-factor model (rFF), as well as two popular composite proxies not examined in any of the three earlier studies (rHL and rDKL). The implied cost of capital proxies are derived from the short-horizon form of the classic dividend discount model, which equates current stock price to a nite series of expected future cash ows and a terminal value, discounted to the present at the cost of equity capital. Alternative estimates arise from models that deal with the terminal value in different ways. Table 2 summarizes the assumptions that underlie the nine unique implied cost of capital proxies and the Value Line forecasts needed to estimate each proxy.12 We follow the estimation procedures employed in prior literature. Specically, we follow the procedures outlined in: (1) BP to compute rDIV, rGLS, rGOR, rOJN, and rPEG; (2) GKS to compute rCT; (3) EM to compute rMPEG and rPEGST; (4) Gode and Mohanram 2003 to compute rGM; (5) Barth, Konchitchki, and Landsman 2010 to compute rFF; (6) Hail and Leuz 2006, 2009 to compute rHL; and (7) Dhaliwahl, Krull, and Li 2007 to compute rDKL. We provide a brief description of each of the proxies below, with an emphasis on key similarities and differences. Table 3 provides details that are not repeated in the text below. rDIV The target price method relies strictly on current stock price and analysts forecasts of dividends and target prices. It employs a short-horizon form of the dividend discount model in which a forecasted terminal value truncates the innite series of future cash ows at the end of Year 5. The only empirical assumption underlying this method is that analysts forecasts of dividends during the forecast horizon and target price at the end of the forecast horizon capture the markets expectation of those values. Beliefs about the evolution

12. As noted in the text, two of the proxies are based on an average of a subset of these, whereas rFF is not an implied cost of capital estimate.

TABLE 2 Summary of forecast assumptions and data requirements for implied cost of capital expected return proxiesa Terminal value Beyond the forecast horizon analysts forecasts of stock price equal the markets expectation. Beyond the forecast horizon zero growth in abnormal earnings. Value Line forecasts

Short-horizon

rDIV

During the forecast horizon, analysts dividend forecasts equal the markets expectation.

rPEG

Dividends per share; current year, one year ahead, and long range. Long range minimum and maximum target price. Earnings per share; one and two year ahead and long range.

rMPEG

Earnings per share; one and two year ahead. Dividends per share; current year.

rPEGST

Analysts earnings forecasts in Years 1 and 2 equal the markets expectation. Zero dividends in Year 1. Year 1 earnings and Year 2 abnormal earnings (specication 2b) are positive. Analysts earnings forecasts in Years 1 and 2 equal the markets expectation. Year 1 earnings and Year 2 abnormal earnings (specication 2b) are positive. Analysts earnings forecasts in Years 1 and 2 equal the markets expectation. Zero dividends in Year 1. Year 1 earnings and Year 2 abnormal earnings (specication 2b) are positive. Beyond the forecast horizon zero growth in abnormal earnings.

(The table is continued on the next page.)

1095

TABLE 2 (Continued) Terminal value Value Line forecasts Dividends per share; current year. Earnings per share; one year ahead and long range

1096

Short-horizon

CAR Vol. 28 No. 4 (Winter 2011) Growth in abnormal earnings is a constant rate for all t. Estimated abnormal earning growth rate equals the markets expectation. Growth rate is less than the cost of equity and exceeds zero. Beyond the forecast horizon, each rms ROE equals its cost of equity capital. Beyond the forecast horizon, rms earn their industry ROE in perpetuity. Beyond the forecast horizon, rms have a 100% dividend payout ratio. Beyond Year 5, abnormal earnings grow at a constant rate, which is assumed to be the ination rate. Book value per share; current year, one year ahead and long range. Earnings per share; one year ahead and long range.

rGOR

rGLS

Dividends per share; current year, one year ahead and long range. Earnings per share; long range. Earnings per share; current year, one year ahead, and long range. Dividends per share; one year ahead and long range. Book value per share; current year, one year ahead, and long range.

rCT

Analysts earnings forecasts in Years 1 and 2 and of dividends in Year 1 equal the markets expectation. Year 1 earnings and Year 2 abnormal earnings (specication 1b) are positive. During the forecast horizon, analysts dividend forecasts equal the markets expectation. During the analysts forecast horizon, analysts forecasts of earnings and book value equal the markets expectation. During the forecast horizon without forecasts, rm ROE fades linearly to industry ROE. Analysts earnings forecasts from Years 1 to 5 equal the markets expectations. Analysts book value forecasts from Years 1 to 4 equal the markets expectation.

Notes:

The assumptions above are stated in general terms and underlie the theoretical form of each model. For empirical purposes, additional assumptions are imposed regarding the length of the forecast horizon.

Two denitions of abnormal earnings are used in the various models. Specication 1 is dened as r)1(eps2 + rdps1 ) Reps1). Specication 2 is dened as r)1(eps2 ) Reps1).

TABLE 3 Formulas for expected return proxies Formula P0 rPEG epsp eps4 =P0 5 rMPEG A A2 eps2 eps1 =P0 ; A dps1 =2P0 rPEGST rOJN p eps2 eps1 =P0 q A A2 eps1 =P0 feps3 eps2 =eps2 eps5 eps4 =eps4 =2 c 1g; 1 rDIV t dpst t p 1

5 P

Variable 1 rDIV 5 P5

rDIV

rPEG rMPEG

rPEGST

Price-earnings-growth ratio (Easton 2004). Modied price-earnings-growth (Easton 2004). Price-earnings-growth ratio (Easton 2004).

rOJN

rGM P0

4 P

rGOR

Economy-wide growth (Ohlson and JuettnerNauroth 2003). Modied economy-wide growth (Gode and Mohanram 2003). Finite horizon (Gordon and Gordon 1997).

A c 1 dps1 =P0 =2 p rGM A A2 eps1 =P0 eps2 eps1 =eps1 c 1; A c 1 dps1 =P0 =2

rGLS

t 1

rCT

1 1 rGOR t dpst rGOR 1 rGOR 4 eps5 t 1 11 P P0 b0 1 rGLS t roet rGLS bt 1 rGLS 1 rGLS 11 1 roe12 rGLS b11 t 51 P P0 b0 1 rCT t roet rCT bt 1 rCT g1 1 rCT 5 roe5 rCT b4 1 g rFF Rft bRMRFi;t RM Rf t bSMBi;t SMBt bHMLi;t HMLt bMOMi;t MOMt The average of rCT, rGLS, rMPEG, and rOJN The average of rCT, rGLS, and rGM after limiting each implied estimate to 0.5

rFF

rHL

rDKL

Fama-French and momentum four-factor model (Barth et al. 2010). Mean implied cost of capital (Hail and Leuz 2006). Mean, adjusted implied cost of capital (Dhaliwal et al. 2007).

1097

1098

of cash ows beyond the forecast horizon are permitted to vary across rms as a function of analysts beliefs embedded in target prices. rPEG, rMPEG, and rPEGST The primary assumption underlying the models used to estimate these proxies is that the market expects zero growth in abnormal earnings beyond the forecast horizon. This assumption simplies the dividend discount model sufciently that it can be solved for expected return directly. The rPEG model also assumes that the market expects dividends in Year t+1 to be zero, whereas rMPEG relaxes this assumption. Following BP we employ long-range earnings forecasts (Year t+5 and Year t+4) in lieu of the short-range earnings forecasts (Year t+2 and Year t+1) to estimate rPEG. Doing so increases our sample size and, we believe, is more in keeping with the assumption regarding future cash ows. In contrast, consistent with many studies (including EM), we employ near-term earnings forecasts (Year t+2 and Year t+1) to estimate rPEGST. Finally, as in EM, estimating rMPEG includes a modication for forecasted dividends. rOJN and rGM Ohlson and Juettner-Nauroth (2005) derive an accounting-based valuation model that imposes a series of assumptions regarding the markets expectations of near term earnings, abnormal earnings, and the rates of short- and long-term growth in abnormal earnings. We employ analysts forecasts to calculate short-term earnings growth rates and, consistent with prior research, set the innite growth in abnormal earnings equal to rf less 3 percent. The sole difference between rOJN and rGM lies in the empirical procedures employed in estimating the model: rGM is estimated with short-term growth in earnings, whereas rOJN is estimated with short- and long-term growth in earnings. rGOR This estimate is derived from a nite specication of the Gordon and Gordon 1997 model, which assumes that beyond the forecast horizon the market expects each rms ROE to revert to its expected return. rCT and rGLS. Claus and Thomas (2001) (hereafter CT) and Gebhardt, Lee, and Swaminathan (2001) (hereafter GLS) employ the residual income model derived from the dividend discount model (Ohlson 1995) to specify prices in terms of forecasted return on equity and expected return. The two approaches deal with the terminal value issue in a different manner, however. CT assume that the market expects abnormal earnings beyond the forecast horizon to grow at a constant rate, which they set equal to the ination rate. GLS assume that the market anticipates ROE will linearly fade to an industry-based ROE 12 years hence, which GLS estimate based on historical industry ROE. rFF Barth et al. (2010) among others (e.g., Kothari, Li, and Short 2009) employ expected return estimates based on the Fama-French four-factor model. The four factors included in the model are a market factor, size factor, book-to-market-factor, and price momentum factor. These factors explain some of the variation in realized returns, but use of this model to estimate Et)1(rt) assumes the factors they employ explain variation in Et)1(rt) and offer a complete representation of the sources of risk for which investors demand ` -vis implied cost of capital approaches, compensation. This is a distinct difference vis-a which do not presuppose the set of priced risks. We estimate the expected annual factor returns (bRMRF(i,t), bSMB(i,t), bHML(i,t), and bMOM(i,t)) by rst calculating each factors

CAR Vol. 28 No. 4 (Winter 2011)

1099

average monthly return over the 60 months prior to month m, and then compounding the resulting average monthly returns over the twelve months prior to the beginning of rm is scal year. rFF is calculated as an annualized predicted return computed using compounded monthly returns, consistent with Barth et al. 2010. rHL and rDKL The nal two proxies we examine are averages of certain other proxies. Specically, Hail and Leuz (2006, 2009) calculate rHL as the mean of rCT, rGLS, rMPEG, and rOJN, whereas Dhaliwal et al. 2007 calculate rDKL as the mean of rCT, rGLS, and rGM, after winsorizing each of these values to a maximum value of 0.5. Comparison of terminal value assumptions The rDIV approach assumes analysts beliefs about innite horizon cash ows accord with market participants beliefs imbedded in stock price. All of the other implied cost of capital approaches assume market participants expectations regarding innite horizon cash ows are consistent with the assumptions imposed by the researcher. Because rDIV does not constrain the behavior of innite growth in expected cash ows to be the same across rms, we expect rDIV to display the greatest cross-sectional variation. Whether this reects variation in the underlying Et)1(rt) however, depends on whether analysts forecasts adequately proxy for the markets expectations. Moreover, as discussed previously, researcher imposed assumptions about the behavior of innite horizon growth in cash ows can create a spurious correlation between the Et)1(rt) estimates produced and growth. Nevertheless, if the researcher-imposed assumptions mirror those of market participants, such correlations need not be spurious. Finally, rFF limits the set of priced risk factors to a predetermined set resulting in an association between rFF and the predetermined set of risk factors by construction. 4. Sample construction, descriptive statistics, and correlations Sample construction For each rm-year, we require: (1) CRSP data to calculate realized returns, (2) Value Line forecasts required to estimate expected return and cash ow news proxies, (3) current stock price, and (4) COMPUSTAT data to calculate our rm-specic risk factor variables. The primary sample for our analysis of the association among the expected returns proxies and rm-specic risk factors consists of 17,904 rm-years drawn from 19842004. The primary sample for our analysis of the association among realized returns and the expected return proxies includes 14,521 rm-years from the same period. Fewer observations appear in our realized return analysis because our cash ow news and expected return news variables are change variables, such that a rm must have sufcient data in adjacent years to be included in the sample employed in this analysis. Descriptive statistics Expected and realized return estimates Panel A of Table 4 provides pooled descriptive statistics pertaining to the twelve Et)1(rt) proxies and realized returns (rREAL). Mean (median) estimates of expected return range from 7.36 (7.27) percent for rGLS to 15.60 (14.50) percent for rFF. The high volatility of realized returns (39.91) compared to the low volatility of expected returns (8.73 at most for rFF) is consistent with variability in the unexpected component of the return swamping variation in the expected return component. This underscores the severity of the power issue that could arise if cash ow and expected return news are ignored.

1100

Panel A: Expected return proxiesa rMPEG 12.36% 11.31 5.92 9.53 13.96 28.61 4.61 17,904 4.34 8.48 12.80 27.54 4.63 17,904 5.42 9.49 13.11 21.20 3.67 17,904 5.89 10.11 14.98 30.54 4.99 17,904 3.92 7.48 10.79 17.59 2.98 17,904 0.98 5.27 9.17 15.39 3.57 17,904 4.98 8.82 12.11 18.66 3.17 17,904 0.91 9.45 20.48 41.25 8.73 15,608 11.23 10.22 11.55% 11.14 13.20% 12.11 9.23% 8.93 7.36% 7.27 10.65% 10.37 15.60% 14.50 rPEGST rOJN rGM rGOR rGLS rCT rFF rHL 10.48% 10.13 5.48 8.62 11.90 19.11 2.92 17,904 rDKL 10.39% 9.99 5.12 8.39 11.91 19.83 3.06 17,904

rREAL

rDIV

rPEG

CAR Vol. 28 No. 4 (Winter 2011) Cash ow news (scaled by forecast) CFN_1A )0.99%* )2.44 )268.00 )36.67 20.38 321.54 1.20 14,521 CFN_TVA 2.43%* 0.00 )57.89 )13.33 16.13 84.21 26.94 14,521 Expected return news EWER_N )0.32** )0.42** )3.01 )0.98 0.50 2.06 1.20 14,521 CFN_TV 0.69%* 0.00 )133.33 )23.26 25.02 134.02 49.64 14,521

15.27% 11.72

15.26% 14.35

11.63% 11.09

CFN_1

FSER_N )0.03** )0.02** )0.72 )0.14 0.09 0.63 0.24 14,521 (The table is continued on the next page.)

)0.76%** )0.16**

TABLE 4 (Continued)

Panel C: Firm-specic risk proxiesd UBETA 0.73 0.64 0.42 0.22 4106.62 909.14 0.55 0.50 DM MKVL BP EXGRW 13.54 12.50

MBETA

Mean Median

1.02 0.98

Notes:

rREAL is the buy and hold realized return computed over the 12 months beginning after the month expected return proxies are estimated. See detailed descriptions and calculations for each proxy in Tables 2 and 3.

CFN_1A and CFN_TVA are analogous to CFN_1 and CFN_TV except for the scaling. The original values are scaled by recent stock price, but the variables labeled A are scaled by the absolute value of the original forecast.

Number of observations for these variables is 17,904. **, * denotes signicance at the 0.01 and 0.05 or better levels, respectively (2-tailed t-test).

1101

1102

Explanatory variables Panel B of Table 4 provides descriptive statistics for our cash ow and expected return news proxies.13 CFN_1 captures current year cash ow news via an earnings surprise variable. CFN_1 has a mean (median) value of )0.76 ()0.16) percent of recent stock price. Both the mean and median values are statistically negative, consistent with analyst optimism. CFN_1A measures the earnings surprise as a percentage of beginning earnings forecast. The mean (median) value indicates an earnings surprise of approximately )1.0 ()2.4) percent of the initial earnings forecast. CFN_TV captures the change in analysts expectations of target price over the 12-month realized return period. The mean change in the target price is signicantly positive (0.69 percent of stock price), while the median value is 0.0. This is consistent with good news on average with respect to innite horizon discounted expected cash ows. The mean (median) value of our economy-wide expected return news proxy is signicantly negative at )0.32 ()0.42) indicating an annual decline in the risk free rate of one-third to almost one-half of one percent. Similarly, the mean (median) value of our rm-specic expected return news proxy is signicantly negative at )0.03 ()0.02). Finally, panel C of Table 4 reports pooled mean and median statistics for our rmspecic risk factors including MBETA, UBETA, DM, MKVL, BP, and EXGRW. These statistics describe a sample similar in average market risk to that of the market portfolio with a mean debt-to-market ratio of 42 percent and a market value of equity that is heavily skewed to larger rms. The average book-to-price ratio of 55 percent is consistent with the relatively healthy rate of 13.54 percent average growth in expected earnings. Correlations Table 5 presents Spearman correlations among sets of variables employed in our regression analyses. We present the mean value of the year-by-year correlations along with the number of years (out of 21) that the annual correlation is signicantly (positive negative). The rst row of panel A, Table 5 presents univariate correlations among realized returns and the twelve proxies for Et)1(rt). The correlations are quite small, ranging from 0.00 (rPEGST and rFF) to 0.09 (rGOR). Four of our Et)1(rt) proxies (rOJN, rGOR, rGLS, and rCT) correlate positively with rREAL in ten or more years but, in four cases (rDIV, rPEG, rPEGST, and rGLS), the proxies correlate negatively with rREAL in ve or more years. Nonetheless, this analysis fails to consider critical controls for cash ow and expected return news as modeled in (2). Except for the correlation between rDIV and rGLS, where the correlation is statistically positive in only 19 years, the pair-wise correlations among the implied cost of capital proxies are statistically positive in all 21 years. rHL and rDKL are highly correlated (q = 0.96) despite being the product of an average of somewhat different subsets of alternative measures, and there is a strong positive correlation among rMPEG, rPEGST, and rGM (q > 0.90). This indicates that these groups of estimates capture essentially the same underlying construct. In contrast, the correlation between rFF and the implied cost of capital estimates is almost as likely to be negative as positive. This suggests that rFF and the implied cost of capital estimates do not capture the same underlying construct. Panel B of Table 5 presents the correlations among the variables we employ in testing Hypothesis 1. Consistent with cash ow news playing an important role in explaining realized returns, the correlation between rREAL and the cash ow news proxies is large

13. We provide cash ow news statistics after scaling each component of cash ows news by recent stock price, the variable included in our empirical analyses. We also provide cash ow news statistics after scaling each component by the absolute value of the relevant forecast at the beginning of the period. We provide these data to convey the economic magnitude of the revisions in cash ows.

TABLE 5 Correlations

Panel A: Correlations among expected return proxies rPEGST rOJN rGM rGOR rGLS rCT rFF rHL rDKL

rDIV

rPEG

rMPEG

rREAL

rDIV

0.04 (6 6)

rPEG

rMPEG

rPEGST

rOJN

0.05 (10 3) 0.56 (21 0) 0.73 (21 0) 0.40 (21 0) 0.29 (21 0)

rGM

0.01 (5 4) 0.44 (21 0) 0.51 (21 0) 0.98 (21 0) 0.91 (21 0) 0.38 (21 0)

rGOR

0.09 (12 2) 0.57 (21 0) 0.60 (21 0) 0.62 (21 0) 0.50 (21 0) 0.68 (21 0) 0.53 (21 0)

rGLS

0.08 (12 5) 0.22 (19 0) 0.22 (21 0) 0.39 (21 0) 0.25 (21 0) 0.36 (21 0) 0.30 (21 0) 0.70 (21 0)

rCT

0.08 (13 3) 0.55 (21 0) 0.56 (21 0) 0.51 (21 0) 0.43 (21 0) 0.63 (21 0) 0.43 (21 0) 0.87 (21 0) 0.54 (21 0)

rFF

0.00 (4 4) )0.04 (2 7) 0.00 (21 6) 0.00 (5 7) 0.00 (7 6) )0.01 (3 3) )0.01 (5 7) 0.04 (8 3) 0.07 (9 2) 0.04 (9 3)

rHL

0.06 (9 2) 0.56 (21 0) 0.64 (21 0) 0.79 (21 0) 0.66 (21 0) 0.70 (21 0) 0.73 (21 0) 0.90 (21 0) 0.72 (21 0) 0.80 (21 0) 0.03 (7 3)

0.05 (9 3) 0.49 (21 0) 0.54 (21 0) 0.86 (21 0) 0.74 (21 0) 0.52 (21 0) 0.80 (21 0) 0.84 (21 0) 0.72 (21 0) 0.74 (21 0) 0.03 (8 3) 0.96 (21 0) (The table is continued on the next page.)

1103

TABLE 5 (Continued)

1104

Panel B: Correlations among realized returns, cash ow news, and expected return news proxies. CFN_TV 0.35 (21 0) 0.46 (21 0) )0.01 (7 8) )0.01 (0 1) )0.01 (0 1) EWER_N FSER_N 0.01 0.00 )0.01 )0.01 (10 4) (4 5) (2 3) (2 4)

CFN_1

0.33 (21 0)

CAR Vol. 28 No. 4 (Winter 2011) rMPEG rPEGST rOJN rGM rGOR rGLS rCT rFF rHL 0.04 (6 1) )0.06 (0 8) )0.01 (0 3) 0.01 (4 4) 0.05 (8 1) )0.06 (0 5) )0.03 (1 6) 0.00 (5 6) )0.04 (0 6) )0.16 (0 19) )0.03 (0 5) 0.01 (4 5) 0.06 (7 1) )0.06 (0 7) )0.03 (0 6) 0.01 (3 3) )0.07 (0 6) )0.13 (0 15) 0.00 (4 3) 0.00 (7 6) )0.04 (0 7) )0.05 (3 10) 0.01 (5 3) 0.01 (7 5) )0.08 (0 6) )0.14 (0 19) )0.03 (1 4) 0.01 (6 4) 0.01 (2 2) 0.00 (2 3) )0.01 (3 3) )0.13 (3 14) )0.01 (2 4) )0.12 (0 13) )0.02 (3 3) 0.01 (6 5)

Panel C: Correlations among expected return and cash ow news and expected return new proxies. rDKL

rDIV

rPEG

CFN_1

CFN_TV

EWER_N

FSER_N

TABLE 5 (Continued)

Panel D: Correlations among expected return and risk proxies. rMPEG rPEGST rOJN rGM rGOR rGLS rCT rFF rHL rDKL

rREAL

rDIV

rPEG

MBETA

UBETA

DM

LMKVL

LBP

EXGRW

)0.05 (8 9) )0.04 (8 10) 0.03 (10 8) 0.02 (7 5) 0.07 (8 3) )0.05 (4 10) 0.08 (11 0) )0.05 (1 7) 0.30 (21 0) )0.25 (0 21) 0.46 (21 0) 0.20 (20 0) 0.27 (21 0) 0.14 (18 0) 0.16 (19 0) )0.29 (0 21) 0.33 (21 0) 0.36 (21 0) 0.10 (12 0) )0.02 (2 6) 0.26 (21 0) )0.24 (0 21) 0.40 (21 0) 0.27 (20 0) 0.03 (6 4) )0.09 (1 13) 0.23 (21 0) )0.12 (1 13) 0.35 (21 0) 0.08 (11 1) 0.19 (14 2) 0.10 (10 4) 0.07 (10 1) 0.06 (10 5) 0.01 (7 5) )0.03 (6 9)

0.18 (17 0) 0.11 (15 0) 0.05 (7 1) )0.16 (1 17) 0.24 (20 0) 0.36 (20 0)

0.31 (21 0) 0.19 (19 0) 0.09 (12 0) )0.29 (0 21) 0.30 (21 0) 0.78 (21 0)

)0.02 (2 4) )0.09 (0 15) 0.20 (21 0) )0.16 (0 18) 0.37 (21 0) 0.45 (21 0)

)0.02 (3 7) )0.15 (0 18) 0.39 (21 0) )0.24 (0 20) 0.63 (21 0) 0.06 (9 3)

)0.16 (0 20) )0.25 (0 21) 0.43 (21 0) )0.17 (0 18) 0.69 (21 0) )0.23 (0 21)

)0.01 (3 4) )0.14 (0 18) 0.39 (21 0) )0.25 (0 20) 0.62 (21 0) 0.18 (17 1)

0.00 (4 4) )0.14 (0 17) 0.39 (21 0) )0.24 (0 20) 0.61 (21 0) 0.09 (13 1)

Notes:

Table values are the mean of year-by-year correlations. Figures in parentheses are the number of years the correlation is signicantly (positive negative) at 5% level in year-by-year correlations. All variables are dened in Tables 1 or 3.

1105

1106

(greater than 0.30) and signicantly positive in all 21 years. We also document a strong positive correlation between our cash ow news proxies (0.46), suggesting that good current period cash ow news tends to be associated with good long-horizon cash ow news. There is little relationship between our macroeconomic expected return news proxy (EWER_N) and our rm-specic expected return news proxy (FSER_N), and the fact that neither is strongly correlated with rREAL supports the conclusion in prior research that cash ow news is the primary driver of realized returns. Finally, the expected return news proxies are not highly correlated with the cash ow news proxies, suggesting that cash ow news is distinct from expected return news. Panel C of Table 5 presents correlations between the Et)1(rt) proxies, and our proxies for cash ow and expected return news. There is a fairly strong negative correlation between CFN_TV and several of the Et)1(rt) proxies (rDIV, rPEG,, rOJN, rGOR, rCT, and rHL), which suggests that terminal values tend to decline when cost of equity capital is high at the beginning of the realized return period. The low correlation among most of the Et)1(rt) proxies and our two proxies for expected return news is reasonable since there is no obvious basis for expected return news to be correlated with the initial level of expected returns. Panel D of Table 5 presents the correlation coefcients among rREAL, our Et)1(rt) proxies, and the rm-specic risk factors. Although we use UBETA in our regression model, we include MBETA in this table for discussion purposes. Six of the proxies (rDIV, rPEG, rPEGST, rMPEG, rGM, and rFF) correlate positively with MBETA in at least 11 years, while one (rGLS) correlates negatively with MBETA in 20 years.14 In contrast, only three of the Et)1(rt) proxies are positively related to UBETA in at least 11 years (rDIV, rPEG, and rPEGST), while six of the proxies (rOJN, rGOR, rGLS, rCT, rHL, and rDKL) are negatively related to UBETA in 11 or more years. All Et)1(rt) proxies other than rDIV and rFF are signicantly positively related to leverage in at least 11 years. This nding, combined with the fact that several of the Et)1(rt) proxies are positively correlated with MBETA but not UBETA, highlights the importance of separating leverage risk and market risk. Except for an unexpected negative association between rGLS and EXGRW ()0.23) and an unexpected positive association between rFF and LMKVL, the remaining correlations among the Et)1(rt) proxies and the risk factors accord with expectations. Despite the counterintuitive nature of the negative association between rGLS and EXGRW, this nding is consistent with prior research (e.g., BP 2005; Gebhardt et al. 2001). The particularly strong correlation between EXGRW and rPEG (q = 0.78) is explained in part by the algebraic relationship between rPEG and growth, and serves to underscore the importance of controlling for growth in our rm-specic risk analysis. Nevertheless, EXGRW is also highly correlated with rDIV (q = 0.36), which does not have an algebraic relationship with growth, which is consistent with growth being an important risk factor in its own right. Finally, the correlation between rREAL and the risk proxies are frequently contrary to expectations if rREAL is viewed as a proxy for expected returns. Taken in their entirety, the results presented in panel D suggest that only rDIV, rPEG, and rPEGST are associated with all of the rm-specic risk factors in a theoretically predictable manner. With respect to rDIV and rPEG this nding is consistent with BP, who study a more limited set of proxies over an earlier time period (19831993). Nonetheless, in light of the association among the risk proxies, it is important to examine the association between the expected return and rm-specic risk proxies in a multivariate setting.

14.

Because MBETA captures both market and leverage risk, documented correlations with MBETA might be due to associations with either risk characteristic.

Expected Returns, Realized Returns, and Firm Risk Characteristics 5. Empirical results Test of Hypothesis 1

1107

Baseline model Panel A of Table 6 provides the results of estimating a baseline realized return model that includes our cash ow and expected return news proxies, but excludes the Et)1(rt) proxy. We estimate this model to assess the incremental explanatory power attributable to the Et)1(rt) proxies. We report the time-series averages of the coefcients from annual crosssectional regressions and t-statistics based on the standard error of the coefcient (Fama and MacBeth 1973). The average R2 of the baseline model is 25.9 percent, indicating that our proxies for cash ow and expected return news capture one-quarter of the variation in rREAL. We document a strong positive relation between realized returns and the cash ow news variables (CFN_1 and CFN_TV), consistent with expectations, as well as results in Voulteenaho 2002. We document a signicant negative relation between rREAL and economy-wide expected return news (EWER_N), but the coefcient on rm-specic expected return news (FSER_N) is insignicant. These ndings are consistent with prior research, which concludes that expected return news is driven by macroeconomic, not rm-specic factors (Vuolteenaho 2002; Campbell and Ammer 1993). Alternatively, CFN_TV might capture rm-specic expected return news, leaving no role for FSER_N. Full regression model Panel B of Table 6 presents the results of estimating twelve specications of the realized return model with the various Et)1(rt) proxies. In all specications the coefcients on CFN_1 and CFN_TV are signicantly positive, the coefcient on EWER_N is signicantly negative, and the coefcient on FSER_N is not statistically different from zero. Thus, adding Et)1(rt) proxies to the model has no effect on the associations between realized returns and cash ow and expected return news. Below the coefcient on the Et)1(rt) proxy, we present t-statistics related to whether the mean proxy coefcient is signicantly positive and whether the mean proxy coefcient is equal to 1. We also present the number of years the proxy coefcient is (1) signicantly positive and not different from one, (2) signicantly positive, and (3) signicantly negative. Except for the coefcients on rPEGST and rFF, the mean coefcients on the Et)1(rt) proxies are signicantly positive, with average values ranging from 0.30 (rGM) to 2.14 (rGOR). The sign of the coefcient is most stable across years for rDIV (signicantly positive in 19 of 21 years). Moreover, in all but one specication (rFF) adding Et)1(rt) to the baseline models improves explanatory power. The models employing rDIV and rGOR show the greatest increase in R2 an increase of 15 percent to almost 30 percent in both cases. For several of the proxies (rDIV, rPEG, rOJN, rGLS, rCT, rDKL, and rDKL) the average coefcient is statistically indistinguishable from the theoretical value of one. rDIV and rOJN perform best in this respect with coefcients not different from one in nine years. Also, if the empirical model is well-specied, the intercept should be zero, which is the case in all but one of the specications (rGOR), providing further support for the appropriateness of our models and proxies. While most of the expected return proxies correlate positively with realized returns, rDIV seems to rise to the forefront in terms of strength of results. The average coefcient on rDIV is signicantly positive, but statistically indistinguishable from 1. It is signicantly positive in the greatest number of years (19) and indistinguishable from one in close to half the years (9). The model estimated with rDIV is also tied for the greatest increase in R2 (15 percent increase).

1108

TABLE 6 Regressions of realized returns on cash ow news and expected return news proxies. Panel A: Model 1: rREALit a0 b1 CFN 1it b2 CFN TVit b3 EWER Nit b4 FSER Nit eit Intercept )0.018 ()0.18) CFN_1 (+) CFN_TV (+) EWER_N ()) FSER_N ()) Avg. Adj. R2 0.008 (0.15) 25.9

Panel B: Model 2: rREALit a0 b1 ERit1 b2 CFN 1it b3 CFN STit b4 CFN TVit b5 EWER Nit b6 FSER Nit eit CFN_1 (+) 0.014 (17.99**) CFN_TV (+) 0.002 (16.22**) EWER_N ()) )0.126 ()1.91*) FSER_N ()) 0.018 (0.44) Avg. Adj. R2 29.7

Intercept rDIV >0 =1 =1 + ) rPEG >0 =1 =1 + ) rMPEG >0 =1 =1 + ) rPEGST >0 =1 =1 + ) rOJN >0 =1 =1 + ) rGM >0 =1 =1 + ) rGOR >0 =1 =1 + ) rGLS >0 =1 =1 + ) rCT >0 =1 =1 + ) )0.131 ()1.13)

ER (+) 0.883 (6.67**) (0.40) {9 19 0} 0.893 (3.50**) (0.69) {7 13 3} 0.434 (2.44**) ()2.77**) {5 10 2} 0.322 (1.54) ()3.05**) {7 10 2} 1.157 (4.46**) (0.60) {9 14 0} 0.304 (1.83*) ()2.68**) {4 9 2} 2.141 (6.99**) (3.33**) {5 17 0} 0.964 (3.23*) (0.92) {4 12 0} 1.581 (5.07**) (1.52) {8 16 0}

)0.104 ()1.34)

0.014 (15.51**)

0.002 (15.09**)

)0.131 ()1.95*)

0.008 (0.17)

27.6

)0.056 ()0.53)

0.014 (16.67**)

0.002 (15.66**)

)0.133 ()1.88*)

0.008 (0.17)

26.7

)0.044 ()0.43)

0.014 (16.29**)

0.002 (15.69**)

)0.137 ()1.95*)

0.002 (0.05)

26.9

)0.124 ()1.44)

0.014 (16.47**)

0.002 (15.42**)

)0.125 ()1.94*)

0.014 (0.29)

27.3

)0.045 ()0.42)

0.014 (16.33**)

0.002 (15.48**)

)0.136 ()1.87*)

0.008 (0.15)

26.6

)0.177 ()2.14**)

0.015 (18.16**)

0.002 (16.14**)

)0.114 ()2.06*)

0.007 (0.17)

29.8

)0.056 ()0.67)

0.015 (17.63**)

0.002 (14.90**)

)0.118 ()1.94*)

0.017 (0.37)

28.0

)0.161 ()1.57)

0.015 (17.68**)

0.002 (16.13**)

)0.122 ()1.90*)

0.016 (0.36)

28.2

TABLE 6 (Continued) CFN_1 (+) 0.014 (15.56**) CFN_TV (+) 0.002 (16.83**) EWER_N ()) )0.111 ()1.89*) FSER_N ()) 0.009 (0.45)

1109

Intercept rFF >0 =1 =1 + ) rHL >0 =1 =1 + ) rDKL >0 =1 =1 + ) Notes: 0.032 (0.45)

ER (+) )0.051 ()0.56) ()6.04**) {1 5 7} 1.430 (4.33**) (0.30) {5 13 0} 1.177 (3.74**) (0.65) {8 12 0}

)0.131 ()1.36)

0.014 (18.17**)

0.002 (16.04**)

)0.117 ()1.95*)

0.014 (0.33)

28.0

)0.107 ()1.08)

0.014 (17.64**)

0.002 (15.92**)

)0.120 ()1.93*)

0.013 (0.29)

27.7

This table includes the time-series averages of the coefcients of the 21 annual cross-sectional regressions (19842004) and t-statistics for whether that mean coefcient is statistically positive negative (>0) using the standard error of the coefcient estimates across the 21 years (Fama and MacBeth 1973). In addition, for the ER coefcients, we include the t-statistic for whether the mean coefcient is equal to one (=1) and the number of times the coefcient in the year-byyear regressions is (signicantly positive and equal to one signicantly positive signicantly negative) (based on a 0.05 p-value) (=1 + )) in {}. **,* denotes signicance at the 0.01, 0.05 level or better (1-tailed t-test). Figures in bold are signicant at the 0.01 level or better. All variables are dened in Tables 1 or 3. Sample size is 14,521.

Our results provide support for the construct validity of all the proxies we examine except rPEGST and rFF. This conclusion is inconsistent with GKS and EM, who document insignicant, and in some cases signicantly negative relationships between the implied cost of capital estimates they examine and realized returns. We investigate the source of the difference in our results in section 7. Although the results in Table 6 provide support for the construct validity of 10 of the Et)1(rt) proxies, rREAL is only one of the other measures that can be used to assess the construct validity of alternative Et)1(rt) proxies. A second set of other measures also useful for this purpose is the set of rm-specic risk factors that theory predicts should be associated with Et)1(rt). The next section presents the results of this analysis. Test of Hypothesis 2 Table 7 presents the results of estimating regression equation 5. The model includes the risk-free rate along with ve risk proxies (UBETA, DM, LMKVL, LBP, and EXGRW). We nd that only rDIV, rPEG, and rPEGST are related as expected to the risk proxies included in the model. rPEG and rPEGST are related to all of the proxies consistent with theory. rDIV is related to all but LMKVL, but this is expected if LMKVL and LBP both serve to capture unmeasured risk (Berk 1995). No other expected return proxy performs as well as rDIV, rPEG or rPEGST with respect to the association with rm-specic risk. In addition, if the empirical model is well-specied the intercept should be zero, which it is in these three specications, and the coefcient on the risk-free rate should be indistinguishable from 1, which it is in the models estimated with rDIV and rPEGST. rREAL is not correlated

CAR Vol. 28 No. 4 (Winter 2011)

1110

TABLE 7 Regression of expected return specications on rm-specic risk factors. Intercept (?) rREAL >0 =1 =1 + ) rDIV >0 =1 =1 + ) rPEG >0 =1 =1 + ) rMPEG >0 =1 =1 + ) rPEGST >0 =1 =1 + ) rOJN >0 =1 =1 + ) rGM >0 =1 =1 + ) rGOR >0 =1 =1 + ) rGLS >0 =1 =1 + ) rCT >0 =1 =1 + ) rFF >0 =1 =1 + ) )111.08 ()1.52) rf (+) 14.45 (1.61) (2.57**) {7 6 0} 1.59 (0.98) (0.49) {6 7 1} 0.29 (0.81) ()3.07**) {6 6 5} 0.61 (2.16**) ()1.85**) {3 2 2} 0.63 (1.55) ()1.40) {4 2 2} 0.45 (2.55**) ()4.16**) {5 0 3} 0.90 (3.28**) ()0.49) {3 0 1} 0.31 (0.97) ()4.18**) {4 6 2} )0.07 ()0.18) ()2.78**) {4 5 2} 0.57 (2.11**) ()2.16*) {7 0 6} )0.74 )1.59 ()2.63**) {4 6 2} UBETA (+) 1.12 (0.34) DM (+) 1.36 (1.43) LMKVL ()) )0.14 ()0.28) LBP (+) 2.40 (2.17**) EXGRW (+) )0.07 ()0.60) Avg. Adj. R2 8.3

)2.34 ()0.19)

1.59 (5.42**)

2.51 (8.36**)

0.27 (7.03**)

22.3

4.92 (1.65)

0.45 (3.30**)

1.60 (9.12**)

0.37 (7.56**)

67.3

7.38 (3.35**)

0.17 (1.19)

2.43 (8.07**)

0.16 (11.99**)

28.5

4.30 (1.19)

1.07 (4.01**)

1.85 (8.21**)

0.21 (13.85**)

30.5

6.40 (4.77*)

)0.91 ()9.71#)

)0.06 ()1.06)

0.10 (1.81#)

1.51 (9.75**)

0.22 (9.62**)

41.2

5.59 (2.35**)

0.08 (0.48)

2.27 (8.51**)

0.21 (12.38**)

28.0

7.73 (3.15**)

)0.07 ()0.70)

2.13 (8.27**)

0.02 (2.46**)

39.3

9.81 (4.21**)

)0.21 ()1.06)

0.56 (7.22**)

0.00 (0.07)

40.8

6.52 (3.17**)

)0.03 ()0.26)

1.11 (12.19**)

0.02 (1.43)

15.6

16.34 (4.07**)

4.07 (2.36**)

2.34 (2.80**)

0.15 (0.23)

0.30 (1.04)

)0.06 ()2.72#)

17.9

TABLE 7 (Continued) Intercept (?) rHL >0 =1 =1 + ) rDKL >0 =1 =1 + ) Notes: 7.41 (4.50**) rf (+) 0.39 (1.70) ()3.55**) {5 1 4} 0.47 (1.89**) ()2.57**) {6 2 5} UBETA (+) )0.17 ()1.87#) DM (+) 0.55 (3.90**) LMKVL ()) )0.02 ()0.37) LBP (+) 1.93 (9.99**) EXGRW (+) 0.08 (7.71**)

1111

7.20 (4.08**)

)0.04 ()0.32)

0.61 (7.02**)

)0.06 ()1.53)

1.98 (10.77**)

0.05 (7.38**)

34.5

This table includes the time-series averages of the coefcients of the 21 annual cross-sectional regressions (19842004) and t-statistics for whether that mean coefcient is statistically positive negative (>0) using the standard error of the coefcient estimates across the 21 years (Fama and MacBeth 1973). In addition, for the rf coefcient, we include the t-statistic for whether the mean coefcient is equal to one (=1) and the number of years the coefcient in the year-by-year regressions is (signicantly positive and equal to one signicantly positive signicantly negative) (based on a 0.05 p-value) (=1 + )) in {}. **, * denotes signicance at the 0.01, 0.05 level or better (1-tailed t-test). # denotes signicant in the wrong direction. Figures in bold are signicant at the 0.01 level or better. All variables are dened in Tables 1 or 3. Sample size is 14,521.

in a reasonable manner with any of the risk factors except for LBP, further supporting the conclusion that rm-level realized returns are not a valid construct for Et)1(rt). In summary, the results of this analysis provide support for the validity of rDIV, rPEG and rPEGST, while the results of the realized return analysis provide support for the construct validity of all of the Et)1(rt) proxies except rPEGST and rFF. Taken together, the two sets of analyses provide support for the validity of rDIV and rPEG alone, since these are the only proxies associated with future realized returns and rm-specic risk in the manner predicted by theory. Amalgamated proxies rHL and rDKL attempt to control for noise in the Et)1(rt) proxy by averaging several proxies, but our analysis suggests that neither is superior to their inputs. Nonetheless, we are sympathetic to concerns regarding noise and the argument that averaging several valid estimates, each individually measured with error, could yield a superior proxy. To address this issue, we focus on rDIV and rPEG because we nd the greatest support for their construct validity. To combine these measures into one proxy we use factor analysis (to isolate the covariance between the two original proxies) and a simple average. Based on results (not tabled) from realized return and risk-based analyses we nd that neither measure yields a proxy that dominates rDIV or rPEG alone. 6. Other issues In this section, we consider three other empirical issues that arise frequently in the literature regarding Et)1(rt) proxies: (1) the impact of analyst forecast bias, (2) the efcacy of realized returns for expected returns after controlling for cash ow news, and (3) substituting realizations for analysts forecasts.

CAR Vol. 28 No. 4 (Winter 2011)

1112

Impact of analysts forecast bias Most Et)1(rt) estimates employ analysts forecasts to proxy for the markets expectations of future cash ows, which gives rise to the concern that deviations between the markets expectations embedded in stock price and analysts forecasts lead to measurement error in the Et)1(rt) proxies. GKS and Hou, van Dijk, and Zhang (2009) express the concern that stale forecasts and or bias in analysts forecasts could lead to measurement error in the Et)1(rt) proxies. It is important to note that error or bias in analysts forecasts relative to reported earnings is not the issue. The issue is whether analysts beliefs are consistent with those of the market at the time expected returns are estimated. In addition, even though systematic bias in analyst forecasts (e.g., optimism relative to the markets beliefs) might lead to biased (e.g., overstated) Et)1(rt) estimates, it need not induce spurious correlations when employing the resulting Et)1(rt) proxy in empirical analyses. For this to occur, measurement error in the Et)1(rt) proxy would need to be systematic and correlated with other variable(s) of interest. Even so, unsystematic deviations between analysts and the markets expectations could create noise and reduce the proxys power. Our earlier analysis provides support for the construct validity of rDIV and rPEG, and accordingly our analysis of this issue focuses on these measures. Because the markets beliefs about future cash ows at the time the proxies are estimated are not observable, we identify rm-years for which we have reason to believe that analysts beliefs might have deviated from those of market participants at the time rDIV and rPEG are estimated. We split our sample into consistent and inconsistent subsamples. Consistent (inconsistent) rm-year observations are those for which the sign of the earnings surprise is the same as (different from) the market reaction to the earnings surprise. We expect the observations in the consistent (inconsistent) subsample to be those with the least (greatest) risk of a deviation between analysts forecasts and the markets expectations. Employing the data in each subsample, we reestimate the realized return model (equation (4)) and expected return model (equation (5)) and present the results in Table 8. Panel A presents the realized return model, estimated with the consistent and inconsistent subsamples. Splitting the sample has little impact on the tenor of our conclusions. The coefcients on the Et)1(rt) proxies continue to exhibit strong positive correlations with rREAL after controlling for cash ow and expected return news in both subsamples. Nevertheless, there is an increase (decrease) in the explanatory power of the model estimated with the subset of consistent (inconsistent) observations. This nding holds for both proxies. The R2 of the rDIV specication estimated using the consistent subsample increases by 14.5 percent from 29.7 percent (Table 6) to 34.0 percent; the R2 of the rPEG specication also increases by 14.5 percent from 27.6 percent (Table 6) to 31.6 percent. In contrast, the R2 of the rDIV specication estimated with the inconsistent subsample decreases by 23 percent to 22.8 percent and the R2 of the rPEG specication decreases by 26 percent to 20.4 percent. These results suggest that deviations between analyst expectations and those of the market do not lead to biased or inconsistent results with respect to the coefcients on the proxies for Et)1(rt), but do reduce the power of the analysis.15 Panel B presents the results of estimating the risk model with the two subsamples. The coefcients on the risk factors are similar for the consistent and inconsistent subsamples, although a couple of risk factors that are statistically signicant in Table 7 lose signicance when the sample is split. In the model estimated with rDIV (rPEG) DM (UBETA) is no longer statistically signicant. Since the ndings are consistent across the

15. The decrease in explanatory power might also be a result of a decreased relation between CFN_1 and realized returns for the inconsistent sample.

TABLE 8 Impact of forecast bias Panel A: Realized return regressions Consistent forecasts (7527 obs.) Intercept rDIV rPEG )0.138 ()1.44) )0.109 ()1.09) ER (+) 1.007 (7.21**) 0.991 (3.80**) CFN_1 (+) 0.016 (17.69**) 0.016 (15.02**) CFN_TV (+) 0.002 (15.75**) 0.002 (14.88**) EWER_N ()) )0.127 ()1.90*) )0.132 ()1.94*) FSER_N ()) 0.007 (0.17) )0.004 ()0.09)

1113

Inconsistent forecasts (6794 obs.) ER (+) 1.030 (4.68**) 1.200 (2.94**) CFN_1 (+) 0.010 (9.88**) 0.009 (9.28**) CFN_TV (+) 0.002 (13.14**) 0.002 (12.77**) EWER_N ()) )0.141 ()2.17**) )0.132 ()2.17*) FSER_N ()) 0.050 (0.44) 0.047 (0.70) Avg. Adj. R2 22.8 20.4

Panel B: Regression of expected return on risk factors Consistent forecasts (7527 obs.) rf (+) 1.59 (1.03) 0.40 (0.97) UBETA (+) 1.03 (4.71**) 0.31 (1.28) DM (+) 0.16 (0.82) 0.32 (5.01**) LMKVL ()) )0.12 ()0.81) )0.13 ()3.28**) LBP (+) 2.40 (7.92**) 1.56 (8.24**) EXGRW (+) 0.32 (9.12**) 0.41 (17.08**) Avg. Adj. R2 22.4 68.4

Inconsistent forecasts (6794 obs.) rf (+) 1.58 (1.02) 0.28 (0.78) UBETA (+) 1.06 (4.75**) 0.30 (1.21) DM (+) 0.18 (0.93) 0.39 (8.18**) LMKVL ()) LBP (+) EXGRW (+) 0.32 (9.18**) 0.43 (14.16**) Avg. Adj. R2 22.5 71.5

1114

TABLE 8 (Continued) Notes: This table includes two robustness tests. Panel A includes the realized return regressions after splitting the sample into observations where the sign of the earnings surprise is consistent with the sign of market response to earnings (consistent) and those where the sign of the earnings surprise is inconsistent with the sign of market response to earnings (inconsistent). We include the time-series averages of the coefcients in the 21 annual cross-sectional regressions (19842004), t-statistics using the standard error of the coefcient estimates across the 21 years (Fama and MacBeth 1973), the number of times the coefcient on the ER proxy is (statistically positive and equal to one signicantly positive signicantly negative) in the year-by-year regressions. Panel B includes the risk model after splitting the sample into observations where the sign of the earnings surprise is consistent with the sign of market response to earnings (consistent) and those where the sign of where the sign of the earnings surprise is inconsistent with the sign of market response to earnings (inconsistent). We include the time-series averages of the coefcients in the 21 annual cross-sectional regressions (19842004), t-statistics using the standard error of the coefcient estimates across the 21 years (Fama and MacBeth 1973). **, * denotes signicance at the 0.01 and 0.05 or better levels, respectively (1-tailed t-test). Figures in bold are signicant at the 0.01 level or better. See Table 3 for detailed denitions of all variables.

subsamples, however, these results are also more consistent with a power issue than biased and inconsistent results. rREAL after controlling for cash ow news Some recent studies use realized returns after controlling for cash ows news to proxy for Et)1(rt) (e.g., Ogneva 2008). We examine the validity of this approach by estimating the expected return model (equation (5)) with rREAL as the dependent variable and augmenting the explanatory variables to control for cash ow and expected return news proxies. Panel A of Table 9 presents the results of estimating this model. Even after controlling for cash ow and expected return news, however, the association between rREAL and many of the rm-specic risk factors contradicts theory. The coefcient on UBETA is not signicant, and the coefcients on LBP and EXGRW are signicant but the wrong sign. As an alternative approach, we adopt a two-stage approach. In the rst stage regression, we estimate the residuals (rRESID) from the baseline realized return model shown in panel A of Table 6. In theory, rRESID should be rREAL purged of the unexpected component of realized returns. That is, rRESID should be a proxy for Et)1(rt). In the second stage regression, we estimate the expected return model (equation (5)) with rRESID as the dependent variable. The results, presented in panel B, mirror those presented in panel A. Taken together, we nd no support for the construct validity of rREAL as a proxy for Et)1(rt) after controlling for cash ow and expected return news. Moreover, since the data needed to estimate the cash ow news proxies is the same data needed to estimate rDIV and rPEG, there is no data advantage to using rREAL as a proxy for Et)1(rt). Accordingly, we see no benet to using realized returns as a proxy for expected returns, with or without controlling for news. Substituting realizations for analysts forecasts When analyst forecasts are unavailable some prior work employs realized values of future earnings and or cash ows in the estimation of implied cost of equity capital. For example, in Chen, Chen, Lobo, and Wang 2011 the authors substitute future realizations of ROE for the markets expectations in estimating rPEG. This can be quite problematic,

CAR Vol. 28 No. 4 (Winter 2011)

Panel A: Controlling for CFN EWER_N ()) )0.077 ()1.13) 0.017 (0.25) 0.04 (5.97**) 0.02 (0.81) )14.66 ()1.06) )0.02 ()3.16**) )0.15 ()9.45#) FSER_N ()) rf (+) UBETA (+) DM (+) LMKVL ()) LBP (+) EXGRW (+) )0.01 ()4.69#) Avg. Adj. R2 35.4

INTERCEPT

CFN_1 (+)

CFN_TV (+)

rREAL

1.54 (1.40)

0.011 (11.94**)

0.002 (15.25**)

Panel B: Cost of capital estimate (= residuals from realized returns) regressed on risk proxies UBETA (+) 0.02 (0.78) 0.04 (5.13**) )0.02 ()3.49**) DM (+) LMKVL ()) LBP (+) )0.14 ()8.93#) EXGRW (+) )0.01 ()7.54#) Avg. Adj. R2 15.4

Intercept (?)

rf (+)

rRESID

1.19 (2.87**)

)13.02 ()2.05#)

Notes:

This table includes two robustness tests. Panel A is based on regressing realized returns on the cash ow and expected return news variables (as controls) and the risk factors included in Table 7. Panel B is based on regressing the residuals from regression of realized returns on the cash ow news and expected return news variables. The values presented are the time-series averages of the coefcients in the 21 annual cross-sectional regressions (19842004), t-statistics using the standard error of the coefcient estimates across the 21 years (Fama and MacBeth 1973). **, * denotes signicance in the predicted direction at the 0.01 and 0.05 or better levels, respectively (1-tailed t-test). # denotes signicant in the wrong direction. Figures in bold are signicant at the 0.01 level or better. See Table 3 for detailed denitions of all variables.

1115

1116

because it leads to systematic error in the estimates that is related to ex post cash ow news. The implied cost of capital estimates are biased upward (downward) for rms with ex post good (bad) cash ow news. Employing these estimates in empirical research might yield biased and inconsistent results if other variables of interest (e.g., growth) vary systematically with rms cash ow news. 7. Reconciliation with prior research Almost all the implied cost of capital proxies we examine are positively correlated with realized returns after controlling for cash ow and expected return news, whereas GKS and EM nd that none of the proxies they examine are positively associated with realized returns. This difference is particularly stark with respect to rPEG, since all three studies examine close variants of this proxy. Our results differ from GKS because their model does not include necessary controls for new information. Consistent with this, we document little or no correlation between rREAL and the expected return proxies in a univariate setting, but, after controlling for cash ow and expected return news, we nd the expected relation. EM also conclude that the GKS results suffer from a severe omitted variable bias. A more complicated issue explains the difference between our results and those of EM. The theoretical specication of our realized return model is the same, although our empirical specications are critically different. We employ the change in the risk-free rate to proxy for macroeconomic expected return news, and the change in market beta to proxy for rm-specic expected return news. EMs proxy for expected return news is a scaled measure of the difference in consecutive implied cost of capital estimates. In the remainder of this section of the paper, we demonstrate that although EMs proxy for expected return news is theoretically defensible, it is empirically problematic because it provokes circularity in the empirical model, which confounds the coefcient on the Et)1(rt) proxy included in the model. All implied cost of capital estimates (ICC) are internal rates of return that equate current stock price (P) to some series of expected future cash ows (CF). As noted earlier, ICCs vary across approaches as different CF assumptions arise from different terminal value assumptions. Nevertheless, by construction, all ICC $ f(CF, P), and therefore, all DICC $ f(DCF, DP). The theoretical specication of the realized return model (i.e., equation (2)) is shown below for convenience. rREAL;t Et1 rt Ncf ;t Nr ;t 6:

Empirically, rREAL,t $ f(DP) and Ncf,t $ f(DCF). In EMs empirical specication Nr,t = DICC $ f(DCF, DP). Consequently, the model EM estimate can be described by the following set of relationships: f DP Et1 rt f DCF f DCF ; DP 7:

EMs proxy for expected return news (DICC) is by construction a function of DCF and DP, which are also included in the model as dependent and explanatory variables, respectively. Stated another way, solving (7) for Et)1(rt) yields:16 Et1 rt f DCF f DCF f DP f DP 8:

The right hand side of (8) implies a product that is close to zero. Expected return is not likely to explain realized returns under this empirical specication. Thus, while it

16. Expected returns are increasing in cash ows (holding price constant) and decreasing in price (holding cash ows constant).

1117

is theoretically defensible to use the change in true Et)1(rt) to capture expected return news, it is empirically problematic to use the change in an Et)1(rt) proxy measured via an implied cost of capital approach for this purpose. The resulting provoked circularity in the empirical model provides no role for Et)1(rt) to contribute to the explanation of rREAL,t, and as a result, any ICC estimate included in the model to proxy for Et)1(rt) will be statistically insignicant, regardless of the validity, or lack thereof, of the ICC estimate employed. Moreover, the circularity we are concerned with not only manifests in no association between rREAL,t and the Et)1(rt) proxies, but in a strong association between rREAL,t and EMs expected return news proxies (i.e., DICC).17 To provide further evidence of the impact of EMs empirical specication for expected return news on the coefcient on the Et)1(rt) proxy, we reestimate our realized return model using EMs cash ow and expected return news proxies (hereafter CFN_EM and ERN_EM, respectively). We estimate the model with the ve implied cost of capital proxies that overlap with the prior work (rPEG, rMPEG, rGM, rCT, and rGLS) plus rDIV, since we nd strong support for the construct validity of the latter proxy. Panel A of Table 10 presents these results. As predicted by our analysis above, and consistent with EMs results, the coefcient on EM_ERN is positive and highly signicant in all specications (t-statistics ranging from 3.86 to 15.56). In addition, except for the coefcients on rCT and rDIV, which are signicantly positive and negative, respectively, the coefcients on the Et)1(rt) proxies are statistically insignicant. Finally, we estimate the regression model using EMs cash ow news proxy, but our measure of expected return news. Because our measures of expected return news are independent of the derivation of the implied cost of capital estimates, they do not provoke circularity in the empirical specication of the model. These results, presented in Table 10, panel B, demonstrate that, once the circularity issue is resolved, the coefcients on the Et)1(rt) proxies are signicantly positive. Finally, it is interesting to note the difference in the R2s of the models estimated in Table 10 panel B versus Table 6. For example, the rDIV model in Table 6 has an R2 of 29.7 percent 68 percent higher than the R2 of 17.7 percent in Table 10, panel B for the rDIV specication. The only difference between these models is the empirical proxy for cash ow news. The former employs our empirical proxy, while the latter employs EMs proxy. The higher R2 achieved with our cash ow news proxy provides evidence of its greater explanatory power. 8. Conclusions Existing literature employs two approaches to assess the validity of alternative proxies for rm-specic cost of equity capital or expected return (Et)1(rt)). One approach relies on the theoretical link between realized returns and Et)1(rt), while the second approach relies on the theoretical relation between Et)1(rt) and priced risk characteristics. Based on results from both approaches we conclude that there is support for the construct validity of two of the Et)1(rt) proxies we examine: rDIV and rPEG. We nd it quite plausible that among the alternatives, rDIV and rPEG consistently demonstrate the greatest degree of construct validity. The primary assumption underlying rDIV is that analysts beliefs regarding short-term cash ows and terminal value concur with

17. The cash ow news proxy EM employ in the estimation of their realized return model differs from the cash ow proxies EM employ in the estimation of ICC. This breaks the cycle of near perfect circularity suggested by our analysis, but merely masks the underlying problem, and further complicates the interpretation of the results. That is, in the absence of this substitution we would expect to observe no association between rREAL,t and the Et)1(rt) proxies, but with this substitution the expected outcome is less clear.

1118

TABLE 10 Regressions of various specications of expected returnb on EM cash ow news and expected return news proxies Panel A: rREALit a0 b1 ER it 1 b2 CFN EM it b3 EM ER N it eit Intercept 0.12 0.15 0.16 0.15 0.16 0.03 0.08 (6.25**) (5.34**) (4.01**) (4.15**) (4.24**) (0.83) (2.79**) ER (+) )0.29 )0.43 )0.31 )0.39 0.76 0.46 ()2.45#) ()1.72) ()1.29) ()1.64) (1.83*) (1.36) CFN_EM (+) 0.50 0.49 0.41 0.45 0.44 0.56 0.57 (4.27**) (10.87**) (11.10**) (8.98**) (8.94**) (16.90**) (16.18**) EM_ERN (+) Avg. Adj. R2 10.4 31.2 25.4 17.3 16.7 26.3 32.9

Panel B: rREALit a0 b1 ERit1 b2 CFN EMit b3 EWER Nit b4 FSER Nit ER (+) CFN_EM (+) 0.30 0.51 0.31 0.48 0.48 0.52 0.49 (6.92**) (14.05**) (6.94**) (13.67**) (13.59**) (16.34**) (13.11**) EWER_N (+) 13.18 12.13 12.75 11.96 12.19 11.31 10.80 (1.94*) (2.02**) (2.01*) (1.94**) (1.93*) (1.92*) (2.04**) FSER_N (+) 0.003 (0.04) 0.01 (0.18) 0.01 (0.12) 0.01 (0.18) 0.01 (0.18) 0.01 (0.02) 0.00 (0.05) Avg. Adj. R2 11.9 17.7 13.8 16.6 16.5 18.0 17.9

Intercept )0.03 )0.09 )0.09 )0.07 )0.06 )0.19 )0.08 ()0.37) ()1.30) ()1.07) ()0.85) ()0.80) ()2.49**) ()1.17)

*, * denotes signicance at the 0.01 and 0.05 or better levels, respectively (1-tailed t-test). # denotes signicant in the wrong direction. Figures in bold are signicant at the 0.01 level or better. See Table 3 for detailed denitions of all variables. t-statistics in the table are based on the time-series averages of the coefcients in the 22 annual cross-sectional regressions 19842004 (Fama and MacBeth 1973).

those of market participants embedded in stock price. This assumption is not unique to our study, and is supported by existing research (Barron, Harris, and Stanford 2005). Importantly, because this is the only researcher assumption imposed on terminal value in the estimation of rDIV, terminal values are free to reect whatever assumptions analysts make with regard to innite horizon cash ows and future discount rates. Accordingly, rDIV is not constrained across rms or industries by researcher-imposed assumptions regarding the behavior of terminal values. Further, the primary researcher-imposed assumption underlying rPEG is that, beyond the forecast horizon, growth in abnormal earnings is zero. It is reasonable to expect this researcher-imposed assumption mirrors the assumption frequently employed by analysts and market participants, since it is commonly taught in nancial statement analysis courses. For example, in their discussion of terminal values, Palepu, Healy, and Bernard (1999) state: But in the face of competition, one would typically not expect a rm to extend its supernormal prots to new additional projects year after year.. .. Each new project would generate cash ows with a present value no greater than the cost of investment the investment would be a zero net present value project. Since the benets of the

CAR Vol. 28 No. 4 (Winter 2011)

1119

project are offset by its costs, it does nothing to enhance the current value of the rm, and the associated growth can be ignored.18 The results of our realized return analysis differ markedly from those documented in prior research. Consistent with EM and our univariate analysis we conclude that the results in GKS are attributable to an omitted variable bias arising from a lack of adequate controls for new information. With respect to EM, we demonstrate that their results are prompted by circularity in their empirical model generated by their empirical approach to measuring expected return news. Finally, we consider several other issues raised in the literature regarding implied cost of capital estimates, including (1) the impact of analysts forecast bias, (2) the efcacy of realized returns for expected returns before and after controlling for cash ow news, (3) the effectiveness of averaging several proxies to produce superior measures, and (4) the substitution of realized values for analysts forecasts of cash ows or earnings. Our evidence suggests that the impact of deviations between analysts expectations and those of the market is limited to potentially less powerful proxies. On the second point, we nd that realized returns are not a reliable proxy for expected returns even after controlling for cash ow news. On the third point we nd that the act of averaging several proxies does not yield an enhanced metric. Finally, we note that substituting realized values for analysts forecasts in the estimation of implied cost of equity capital yields estimates that are systematically biased upward (downward) for rms with ex post good (bad) cash ow news, which could yield biased and inconsistent results if the resulting measurement error is correlated with other variables of interest. In conclusion, we recommend that researchers requiring a valid Et)1(rt) proxy employ either rDIV or rPEG estimated with analysts forecasts and we caution against the use of realized returns with or without controlling for cash ow news to proxy for Et)1(rt). We advocate that researchers assess the validity of any new Et)1(rt) proxies by demonstrating a consistent and predictable association between the proxy and future realized returns, as well as established risk measures. Finally, we note that the primary difference between rDIV and rPEG is that rDIV effectively allows the terminal value assumption to vary across rms, while rPEG imposes an assumption of zero growth in abnormal earnings beyond the forecast horizon on all rms regardless of their circumstances. This suggests that rDIV might be superior even to rPEG for rms with nonzero growth in abnormal earnings beyond the forecast horizon. We leave an investigation of this supposition for future research.

18.

Palepu et al. (1999, 126). This is one example, but similar instruction can be found in almost any nancial statement analysis text.

1120

Et)1(rt) proxy rDIV rPEG rMPEG rPEGST rOJN rGM rGOR rGLS rCT rFF rHL rDKL rdagr rPE Notes: rdagr is the expected return estimate imputed from Eastons 2004 implementation of the Ohlson and Juettner-Nauroth 2005 model. rPE is the expected return estimate imputed from the price to forward earnings model. All other expected return estimates are dened in Tables 2 and 3. Guay, Kothari and Shu (2005) Easton and Monahan (2005) Botosan and Plumlee (2005) x x x x x x x x Current study x x x x x x x x x x x x

x x x x

x x x

x x

References

Ashbaugh-Skaife, H., D. W. Collins, W. R. Kinney Jr, and R. LaFond. 2009. The effect of SOX internal control deciencies on rm risk and cost of equity. Journal of Accounting Research 47 (1): 143. Barron, O., D. Harris, and M. Stanford. 2005. Evidence that investors trade on private event-period information around earnings announcements. The Accounting Review 80 (2): 40321. Barth, M., Y. Konchitchki, and W. Landsman. 2010. Cost of capital and earnings transparency. Working paper, Stanford University. Beaver, W., P. Kettler, and M. Scholes. 1970. The association between market determined and accounting determined risk measures. The Accounting Review 45 (4): 65481. Berk, J. 1995. A critique of size-related anomalies. Review of Financial Studies 8 (2): 27586. Botosan, C., and Plumlee, M. 2002. A re-examination of disclosure levels and expected cost of equity capital. Journal of Accounting Research 40 (1): 2140. Botosan, C., and Plumlee, M. 2005. Assessing alternative proxies for the expected risk premium. The Accounting Review 80 (1): 2153. Campbell, J. 1991. A variance decomposition for stock returns. The Economic Journal 101 (405): 15779. Campbell, J., and J. Ammer. 1993. What moves the stock and bond markets? A variance decomposition for long-term asset returns. The Journal of Finance 48 (1): 337. Carmines, E. G., and R. A. Zellner. 1979. Reliability and validity assessment, Thousand Oaks, CA: Sage. Chan, L., and J. Lakonishok. 1993. Are the reports of betas death premature? Journal of Portfolio Management 19 (4): 5162. Chen, H., J. Z. Chen, G. J. Lobo, and Y. Wang. 2011. Effects of audit quality on earnings management and cost of equity capital: Evidence from China. Contemporary Accounting Research. Forthcoming.

1121

Chung, K. 1989. Debt and risk: A technical note. Journal of Business Finance and Accounting 16 (5): 71927. Claus, J., and J. Thomas. 2001. Equity premia as low as three percent? Evidence from analysts earnings forecasts for domestic and international stock markets. The Journal of Finance 56 (5): 162966. Dhaliwal, D., L. Krull, and O. Li. 2007. Did the 2003 Tax Act reduce the cost of equity capital? Journal of Accounting and Economics 43 (1): 12150. Easley, D., S. Hvidkjaer, and M. OHara. 2002. Is information risk a determinant of asset returns? The Journal of Finance 57 (5): 2185221. Easton, P. 2004. PE ratios, PEG ratios, and estimating the implied expected rate of return on equity capital. The Accounting Review 79 (1): 7396. Easton, P., and S. Monahan. 2005. An evaluation of the reliability of accounting based measures of expected returns: A measurement error perspective. The Accounting Review 80 (2): 50138. Elton, J. 1999. Expected return, realized return, and asset pricing tests. The Journal of Finance 54 (4): 1199220. Fama, E., and J. MacBeth. 1973. Risk, return and equilibrium: Empirical tests. Journal of Political Economy 81 (3): 60736. Gebhardt, W., C. Lee, and B. Swaminathan. 2001. Towards an ex-ante cost of capital. Journal of Accounting Research 39 (1): 13576. Gode, D., and P. Mohanram. 2003. Inferring the cost of capital using the Ohlson-Juettner model. Review of Accounting Studies 8 (3): 399431. Gordon, J., and M. Gordon. 1997. The nite horizon expected return model. Financial Analysts Journal 53 (3): 5261. Guay, W., S. P. Kothari, and S. Shu. 2005. Properties of implied cost of capital using analysts forecasts. Working paper, Massachusetts Institute of Technology. Hail, L., and C. Leuz. 2006. International differences in the cost of equity capital: Do legal institutions and securities regulation matter? Journal of Accounting Research 44 (3): 485531. Hail, L., and C. Leuz. 2009. Cost of capital effects and changes in growth expectations around US cross-listings. Journal of Financial Economics 93 (3): 42854. Hamada, R. 1972. The effect of the rms capital structure on the systematic risk of common stocks. The Journal of Finance 27 (2): 43552. Hou, K., M. A. van Dijk, and Y. Zhang. 2009. The implied cost of capital: A new approach. Working paper, The Ohio State University. Kester, W., W. Fruhan, T. Piper, and R. Ruback. 1997. Case problems in nance. New York: McGraw-Hill Irwin. Kothari, S. P., X. Li, and J. Short. 2009. The effect of disclosures by management, analysts, and nancial press on cost of capital, return volatility, and analyst forecasts: A study using content analysis. The Accounting Review 82 (5): 125597. Lambert, R. A. 2009. Discussion of On the relation between expected returns and implied cost of capital. Review of Accounting Studies 14 (23): 26068. Lee, C., E. So, and C. Wang. 2010. Evaluating implied cost of capital estimates. Working paper, Stanford University. Lintner, J. 1965. The valuation of risk assets and the selection of risky investments in stock portfolios and capital budgets. Review of Economics and Statistics 47 (1): 1337. McInnis, J. 2010. Earnings smoothness, average returns, and implied cost of equity capital. The Accounting Review 85 (1): 31541. Modigliani, F., and M. Miller. 1958. The expected cost of equity capital, corporation nance, and the theory of investment. American Economic Review 48 (June): 26197. Monahan, S., and P. Easton. 2010. Evaluating accounting-based measures of expected returns: Easton and Monahan and Botosan and Plumlee redux. Working paper, Notre Dame University.

1122

Mossin, J. 1966. Equilibrium in a capital asset market. Econometrica 34 (4): 76883. Ogneva, M. 2008. Accrual quality and expected returns: The importance of controlling for cash ow shocks. Working paper, Stanford University. Ogneva, M., K. R. Subramanyam, and K. Raghunandan. 2007. Internal control weakness and cost of equity: Evidence from SOX section 404 disclosures. The Accounting Review 82 (5): 125597. Ohlson, J. 1995. Earnings, book value, and dividends in security valuation. Contemporary Accounting Research 11 (2): 66187. Ohlson, J., and B. Juettner-Nauroth. 2005. Expected EPS and EPS growth as determinants of value. Review of Accounting Studies 10 (23): 34965. Palepu, K., P. Healy, and V. Bernard. 1999. Business analysis and valuation: Using nancial statements, 2nd ed. Mason, OH: Thomson South-Western. Pastor, L., M. Sinha, and B. Swaminathan. 2008. Estimating the intertemporal risk-return tradeoff using the implied cost of capital. The Journal of Finance 63 (6): 285997. Ross, S. 1976. The arbitrage theory of capital asset pricing. Journal of Economic Theory 13 (3): 34160. Sharpe, W. 1964. Capital asset price: A theory of market equilibrium under conditions of risk. The Journal of Finance 19 (3): 42542. Vuolteenaho, T. 2002. What drives rm-level stock returns? The Journal of Finance 57 (1): 23364.

## Viel mehr als nur Dokumente.

Entdecken, was Scribd alles zu bieten hat, inklusive Bücher und Hörbücher von großen Verlagen.

Jederzeit kündbar.