Costs of Capital and Earnings Attributes

Jennifer Francis* (Duke University)

Ryan LaFond

(University of Wisconsin)

Per Olsson (Duke University)

Katherine Schipper

(Financial Accounting Standards Board) We examine the relation between the cost of equity capital and seven attributes of earnings: quality, persistence, predictability, smoothness, value relevance, timeliness and conservatism. We refer to the first four attributes as accounting-based because measures of these constructs are typically based on accounting information only. We refer to the last three attributes as market-based because proxies for these constructs are typically based on relations between market data and accounting data. Our results show that firms with the most favorable values of each attribute, viewed individually, enjoy significantly lower costs of capital than firms with the least favorable values. The largest cost of capital effects are found for the accounting-based attributes; within this set, earnings quality has the strongest effects. Among the market-based attributes, value relevance dominates timeliness and conservatism. Considering all attributes together, we find that investors consistently price earnings quality and earnings persistence, and to a lesser extent, value relevance.

Draft: May 2003

Corresponding author: Fuqua School of Business, Duke University, Durham, NC 27708. Email address, jfrancis@duke.edu. This research was supported by the Fuqua School of Business, Duke University and the University of Wisconsin. We are grateful to Alon Brav for access to estimates of ex ante costs of equity capital. We appreciate comments from Moshe Bareket, Alon Brav, and DJ Nanda, and workshop participants at Duke University. The views expressed in this paper are those of the authors and do not represent positions of the Financial Accounting Standards Board. Official positions of the Financial Accounting Standards Board are arrived at only after extensive due process and deliberation.

Costs of Capital and Earnings Attributes 1. Introduction

We investigate the association between attributes of accounting earnings and investors’ resource

allocation decisions, using the cost of equity capital as a summary indicator of those decisions.

Specifically, we analyze the extent to which firms with favorable values of seven characteristics,

suggested by prior research as capturing desirable features of earnings, enjoy a lower cost of capital. We

start from the premise of a relation between the cost of capital and properties of firm-specific information,

and from the presumption that earnings is a premier source of firm-specific information.1 The properties

we consider are attributes used in previous accounting research to characterize earnings: quality,

persistence, predictability, smoothness, value relevance, timeliness and conservatism. Our intent is to

shed light on which earnings attributes appear to matter most to investors, as evidenced by the strength of

association between the attributes and our measure of their resource allocation decisions.

Based on the way prior research has operationalized the earnings attributes we consider, we refer

to quality, persistence, predictability and smoothness as “accounting-based” because they are measured

using accounting information only. Prior studies measure earnings quality using either the mapping of

current accruals into cash flows or some measure of abnormal accruals (e.g., Dechow and Dichev [2002];

Francis, LaFond, Olsson and Schipper [2002]); measures of earnings persistence typically rely on the

estimated slope coefficient in a regression of current earnings on lagged earnings (e.g., Lev [1983]);

measures of earnings predictability focus on the prediction errors from a time-series earnings model (e.g.,

Lipe [1990]); and smoothness measures are based on the volatility of earnings relative to some

benchmark, such as cash flows (e.g., Leutz, Nanda and Wisocki [2003]; Thomas and Zhang [2002]).

Also following previous research, we refer to value relevance, timeliness and conservatism as

“market-based” earnings attributes because measures of these constructs are based on the estimated

relation between accounting earnings and market prices or returns. For example, many prior studies 1 Previous analytical and empirical research which demonstrates this relation includes Easley and O’Hara [2001]; Easley, Hvidkjær and O’Hara [2002]; Francis, LaFond, Olsson and Schipper [2002]; and Botosan [1997].

1

measure value relevance as the R2 from a regression of annual returns on annual earnings (e.g., Collins,

Maydew and Weiss [1997]; Francis and Schipper [1999]). Reverse regressions of earnings on variables

capturing positive stock returns and negative stock returns provide measures of timeliness (the reverse

regression R2; see, for example, Bushman, Chen, Engel and Smith [2003]; Ball, Kothari and Robin

[2000]) and conservatism (the ratio of the reverse regression coefficient on negative returns to the

coefficient on positive returns; see, for example, Basu [1997]; Pope and Walker [1999]).

Our investigations are motivated by previous research which uses earnings attributes to describe

desirable characteristics of reporting systems, with the implication that those involved in the financial

reporting process (standard setters, preparers, auditors, enforcement authorities) should strive for earnings

with certain attributes.2 Assuming that earnings attributes are desirable to the extent they result in a

discernible capital market advantage conferred by investors, we choose as our measure of this advantage

the equity cost of capital. That is, we interpret prior research as suggesting predictable associations

between earnings attributes and measures of the cost of equity, for example, an association between lower

costs of equity and more value relevant earnings, controlling for known risk factors.

We investigate these associations along three dimensions. First, we test whether earnings

attributes identified in prior research as desirable or beneficial are individually associated with a lower

cost of capital. Second, we analyze whether measures of these individual associations vary across

earnings attributes, that is, whether the capital market benefits associated with some earnings attributes

exceed the benefits associated with other attributes. Finally, we test for conditional effects of earnings

attributes, based on assessments of the association between the cost of equity capital and each attribute,

2 For example, Lev and Zarowin [1999] attribute their finding of decreased value relevance of financial information to shortcomings in the financial reporting model, and imply that the changes they recommend would result in increased value relevance. As another example, Joos and Lang [1994] interpret their finding of no increase in German and French firms’ R2 values and coefficient estimates (from regressions of 12-month returns on the level and change in annual earnings) in the post European Union directive period, as suggesting that the directives did little to improve the value relevance of financial reporting in these countries. The implication from both studies is that value relevance is a beneficial attribute of earnings, which standard setters and policy makers should consider in setting financial reporting standards.

2

conditional on the others. These conditional tests are aimed at identifying whether earnings attributes

have distinct cost of equity effects or whether one or more attributes subsumes the others.

We use two complementary approaches to assessing the association between cost of capital

estimates and earnings attributes. The first approach uses ex ante cost of equity estimates derived from

Value Line (VL) analyst forecasts. Using cross-sectional regression estimates for a sample of over

85,000 firm-quarters over 1975-2001 (representing 1,865 distinct firms), we find that each earnings

attribute, considered separately, is significantly associated with the cost of equity capital, controlling for

known risk proxies (beta, size and the book to market ratio). Specifically, firms with the most favorable

values of each attribute have significantly lower costs of equity than firms with the least favorable values.

The largest individual effect is a 396 basis point (bp) differential between the best and worst earnings

quality deciles and the smallest effect is a 34 bp differential between the most and least conservative

earnings deciles. In between, earnings persistence has a 216 bp spread, earnings smoothness and earnings

value relevance have 114-120 bp spreads, and earnings timeliness and earnings predictability have 56-71

bp spreads. In conditional tests which consider all seven earnings attributes jointly, smoothness,

timeliness and conservatism are no longer associated with the cost of equity; predictability is inversely

associated with the cost of equity (conditional on the other earnings attributes, firms with more

predictable earnings have higher, not lower, costs of equity), and three attributes continue to be strongly

positively associated with the cost of equity: quality (416 bp spread), persistence (158 bp spread) and

value relevance (71 bp spread).

The second approach uses realized returns as the benchmark for assessing the cost of capital.

Specifically, we use time-series asset pricing regressions (based on the conventional three-factor model)

to examine whether mimicking factors for the earnings attributes are priced incrementally to known risk

factors. Results of these tests are broadly similar to those of the cross-sectional tests, in that all earnings

attributes are individually associated with the cost of capital in the predicted direction. In conditional

tests, we find that earnings quality has the largest cost of capital effect of all of the earnings attributes we

consider and persistence has statistically positive but smaller effects. In contrast to results obtained from

3

cross-sectional regressions, value relevance (as well as the other market-based attributes) is not reliably

associated with the cost of capital in the time-series analyses, once we include all seven earnings

attributes.

Each approach has its advantages and limitations. The main benefit of the cross-sectional

approach lies in its focus on ex ante costs of equity capital, as opposed to cost of equity measures based

on realized returns.3 The limitation of the cross-sectional tests is that the sample is restricted to VL firms

with the data required by our tests. To show that our results are not driven either by the use of ex ante

measures or by features of the sample, we also examine measures based on realized returns; these time-

series tests provide three kinds of insights. First, they provide an independent (from the cross-sectional

tests) source of evidence that earnings attributes are incorporated in investors’ resource allocation

decisions, as reflected in the cost of equity capital. Second, they demonstrate consistency between the

analyst-based, ex ante cost of capital estimates and investor-based, ex post cost of capital estimates.

Third, the use of factor mimicking portfolios, rather than the attributes themselves, supports the

generalizability of the results because it allows us to examine the cost of capital effects of earnings

attributes for all firms with at least 24 monthly returns over 1975-2001 (18,786 firms).

The earnings attributes we consider are jointly determined by management’s reporting decisions

and by innate features of firms’ operating environments. Therefore, the cost of capital effects we

document also reflect the joint influences of reporting decisions and operating environments. While it is

beyond the scope of this paper to fully separate the cost of capital effects associated with reporting

decisions from those associated with intrinsic or innate firm attributes, we do provide an exploratory

analysis of the incremental effects of earnings attributes after controlling for the innate volatility of firms’

operating environments, as proxied by the standard deviation of cash flows. We find that innate volatility

explains a significant portion of the variation in the cost of equity capital, and that earnings quality

continues to have a statistically reliable, albeit reduced, effect as well.

3 While it is common in the finance and accounting literatures to use realized returns to proxy for expected returns, a growing literature points to the importance of validating such ex post measures with ex ante measures (see, for example, Elton [1999]; Brav, Lehavy and Michaely [2002]; and Gebhardt, Lee and Swaminathan [2001]).

4

We believe our study makes two kinds of contributions. First, we extend previous research which

has considered desirable attributes of earnings one or a few at a time; we consider seven widely used

attributes both individually and conditional on all the others. Second, because we choose a specific

measure of the benefit to be gained from desirable earnings attributes, we are able to provide evidence on

which earnings attributes investors emphasize in their capital allocation decisions, as summarized in the

cost of equity capital. We also estimate the magnitude of the benefits conferred by each attribute.

We interpret our results as suggesting that favorable accounting-based earnings attributes,

particularly high earnings quality and earnings persistence, confer a greater capital market advantage than

do favorable market-based attributes. One implication is that accounting standard setters, who use

decision usefulness as the benchmark to calibrate standard setting decisions, need not be as concerned as

previous research might suggest about the apparent declining value relevance of earnings, or about the

ability of earnings to impound bad news more quickly than good news.

The rest of the paper proceeds as follows. The next section frames our research questions in the

context of the relevant literature. Section 3 describes the sample, the data and the construction of the test

variables. Section 4 reports the results of the cross-sectional tests, and section 5 reports the results of the

time-series tests. Section 6 reports the results of several additional tests, and section 7 summarizes and

concludes.

2. Background, Motivation and Description of Earnings Attributes

Academic accounting research, as well as practice-oriented studies and textbooks, describes

desirable attributes of accounting earnings. For example, research examining the value relevance of a

particular income element or accounting disclosure presumes that value relevance is (or should be) a

sought-after attribute. However, prior research tends to focus on one or two earnings attributes at a time,

and to use differing (sometimes unstated) reference constructs for capturing desirability. Our analysis

uses research which demonstrates capital market consequences of firm-specific information to support the

choice of cost of equity capital as a reference construct against which to calibrate earnings attributes.

5

That is, we consider the cost of equity to be a key investor-driven indicator of the consequences of

differences across firms in the attributes of earnings. The strength of the association between a given

earnings attribute and the cost of capital is a quantitative measure of the desirability of that attribute from

the perspective of capital allocation decisions of investors as summarized in the cost of equity capital.

2.1 Research on capital market consequences of firm-specific information

This stream of research, which includes accounting and finance asset pricing studies,

demonstrates analytically and empirically that firm-specific information factors affect the cost of capital

(Easley and O’Hara [2001]; Easley, Hvidkjær and O’Hara [2002]; Francis, LaFond, Olsson and Schipper

[2002]; and Botosan [1997]). The information factors considered vary across the studies. For example,

Easley and O’Hara focus on information risk faced by relatively uninformed investors because privately

informed investors are better able to shift their portfolio weights to take advantage of new information,

which Easley, Hvidkjær and O’Hara operationalize using PIN (probability of informed trading) scores;

Francis et al. focus on earnings quality (measured as abnormal accruals and, separately, the strength of

mapping of current accruals into cash flows); and Botosan focuses on disclosure scores based on the

quantity of annual report information and analyst perceptions of disclosures, as captured by AIMR scores.

In an international context, Bhattacharya, Daouk and Welker [2003] examine the association between

country-level measures of the average cost of equity and earnings opacity; the latter is defined as a

composite measure of earnings aggressiveness, loss avoidance, and smoothness. Each of these studies

predicts and finds a relation between the information factor they consider and the cost of capital.4 Our

research is premised on the existence of this relation.5

4 Bhattacharya et al’s results, based on a sample of 34 countries over 1986-1998, are sensitive to the cost of equity proxy: earnings aggressiveness is positively associated with their dividend-based cost of equity estimates, while loss avoidance is positively associated with the international pricing model-based cost of equity estimates; in neither case does a country’s earnings smoothness have a measurable effect on its cost of equity. 5 We frame our investigation of earnings attributes from a rational asset pricing perspective, which views information factors as risk factors (see Easley and O’Hara [2001] for a formal rational equilibrium model of priced information risk). An alternative perspective views pricing effects of information factors as anomalies that reflect investor irrationality. Because it is not possible to distinguish between these perspectives (Brav and Heaton [2002]) and because this distinction does not bear on our research question or our design, we do not take a stance on whether the cost of capital effects we document reflect rational pricing or some form of investor irrationality.

6

2.2 Earnings attributes

We build on academic and practitioner studies that propose specific earnings attributes as

desirable outcomes of the financial reporting process. We organize our discussion of this research around

the seven earnings attributes we consider. Descriptions of how we measure each attribute are contained

in section 3.

Earnings quality. Several approaches to assessing earnings quality take cash as the reference

construct; earnings quality increases as earnings are closer to cash (e.g., Penman [2001, p. 611]).

Analysts (e.g., Harris, Huh and Fairfield [2000]) sometimes point to this attribute as particularly

desirable. Dechow and Dichev [2002] propose and test a measure of earnings quality that captures the

mapping of current accruals into last-period, current-period and next-period cash flows, and Francis et al.

[2002] demonstrate that this measure of earnings quality is associated with measures of the cost of both

debt and equity capital. We use the Dechow-Dichev measure to capture the quality attribute of earnings.

Persistence is viewed as a measure of earnings sustainability; persistent earnings are desirable

because they are recurring (e.g., Penman and Zhang [2002]). Analysts sometimes focus on sustainable or

recurring earnings (see, for example, chapter 6 of AICPA [1994]). Greater earnings persistence has been

shown to be associated with larger slope coefficients relating returns to earnings (Kormendi and Lipe

[1987]; Easton and Zmijewski [1989]; Collins and Kothari [1989]).

Predictability. Following Lipe [1990], we define this construct as the ability of earnings to

predict itself. Predictability is an element of relevance in the FASB’s Conceptual Framework, and is

therefore a desirable earnings attribute from the perspective of standard setters. Predictability is also

valued by analysts (see, for example, the AIMR’s description of the distinction between financial

reporting and financial analysis; AIMR [1993]), and is an essential component of valuation (see, for

example, Lee [1999] for a discussion).

Smoothness. Discussions of the benefits of smooth earnings include Ronen and Sadan [1981],

and Chaney and Lewis [1995]. Arguments that smoothness is a desirable earnings attribute derive from

the view that managers use their private information about future income to smooth out transitory

7

fluctuations and thereby achieve a more representative, hence more useful, reported earnings number.

We follow Leuz, Nanda and Wisocki [2003] in using cash flows as the reference construct for

unsmoothed earnings, and measure smoothness as the ratio of income variability to cash flow variability.

Value relevance. This construct is often measured as the ability of earnings to explain variation

in returns, where greater explanatory power is viewed as desirable. The reference construct is therefore

stock prices or stock returns. One stream of this research interprets value relevance as a measure of

usefulness (e.g., Collins, Maydew and Weiss [1997]; Francis and Schipper [1999]). This interpretation

rests on the view that value relevance measures capture combined relevance and reliability, two key

concepts in the FASB’s Conceptual Framework (for an extended discussion, see Barth, Beaver and

Landsman [2001]).

Timeliness and conservatism. These two attributes derive from the view that accounting earnings

is intended to measure economic income, defined as changes in market value of equity (see, for example,

Ball, Kothari and Robin [2000]). Conservatism differs from timeliness in that it reflects the ability of

accounting earnings to differentially reflect economic losses (measured as negative stock returns) and

economic gains (measured as positive stock returns). The reference construct for both timeliness and

conservatism is therefore stock returns, but the two constructs differ in that timeliness does not

distinguish between positive and negative returns and conservatism focuses on the latter. Both measures

are based on reverse regressions of earnings on returns; timeliness is the explanatory power of the

regression and conservatism is the ratio of slope coefficients on negative returns to slope coefficients on

positive returns.6 Combined timeliness and conservatism are sometimes described as “transparency,” a

desirable attribute of accounting earnings (see, for example, Ball et al.).

6 In sensitivity tests (section 6), we consider an accounting-based measure of conservatism -- Penman and Zhang’s [2002] conservatism measure, C-score.

8

3. Sample, Data and Variable Measurement

Our sample covers the 27-year period, t=1975-2001. We calculate our proxies for the earnings

attributes over rolling 10-year windows; a firm is included in the year t sample if the necessary data to

calculate the attribute proxies are available in years t-9 to t. To mitigate concerns that differences in

sample composition drive results comparing accounting-based and market-based attributes, we further

require that data on all seven attributes are available for each firm-year. Table 1 shows that the number of

firms meeting these requirements (the “Full Sample”) ranges from 678 to 1,997 per year, for an average

of 1,471 firms per year. In total, the Full Sample contains 3,917 distinct firms. Table 1 also shows that

the Full Sample represents an average of 53% of the total CRSP market capitalization.7

Our cross-sectional tests also require data on the cost of equity capital. We use Brav, Lehavy and

Michaely’s [2002] ex ante measures, which are derived from data provided in Value Line (VL) reports.8

We have ex ante cost of equity measures, , for each quarter, for VL firms over 1975-2001. Over

this period, Value Line followed about 1,700 firms; our sample size is reduced by the requirement that a

firm have data on CofC and on the seven attributes. Table 1 shows that the number of firms meeting

these requirements (the “VL Sample”) ranges between 524 and 1,022 firms per year, with an average of

790 firms per year. In total, the VL Sample contains 1,865 distinct firms. In terms of representativeness,

Table 1 shows that the VL Sample covers, on average, 47% of the total CRSP market capitalization.

CofC

Table 2 reports descriptive information about the implied cost of capital estimates. Because VL’s

13-week forecast cycle does not conform to either calendar or fiscal quarters, we assign the VL

observations on CofC to each calendar quarter using the date of the VL report containing the forecast

7 Our measure of the total CRSP market capitalization includes all listed securities with share price and shares outstanding data available in December of each year. 8 Brav, Lehavy and Michaely [2002] determine ex ante (implied) cost of equity estimates using the VL analyst’s 4-year out target price (TP), as well as his forecast of next period dividends (DIV) and dividend growth (g). Assuming that interim dividends are reinvested at the firm cost of capital (CofC), Brav et al. arrive at the following expression

for the ex ante expected return:

4 4

4

(1 ) (1 )

)

CofC gDIVCofC gTPCofC

P P(1

+ − + − + = + where P = stock price nine days

prior to the date of the VL report. The value of CofC that satisfies the equality is the estimate of the ex ante cost of equity capital.

9

information, e.g., a February VL report containing the necessary data to calculate the implied CofC is

assigned to the first calendar quarter. Table 2 shows the yearly means (calculated across firm-quarters)

range between 12.41% (in 1997) and 33.19% (in 1975); the pooled sample average (median) is 20.83%

(20.22%), with a standard deviation of 7.76%. These data are broadly similar to cost of capital data used

in other studies. For example, Botosan [1997] reports mean (median) cost of equity estimates for a

sample of 122 VL firms in 1991 of 20.1% (19.0%); for our VL sample of 775 firms in 1991, the mean

(median) CofC is 18.3% (17.1%). Botosan and Plumlee [2002] report mean (median) cost of capital

estimates of 16.5% (15.6%) for a sample of 668 firms over 1986-1996; for our VL sample, the mean

(median) CofC for this same period is 16.1% (15.3%).

Our analyses require measures of the seven earnings attributes that we study. Our construction of

these measures is based on prior research, as is our classification of the measures as accounting-based

(that is, measured by reference to accounting information only) and market-based (measured by reference

to both stock returns and accounting information). The results of our analyses are not necessarily

generalizable to alternative measures of the attributes (e.g., a measure of earnings persistence based on

market information, or a measure of value relevance based on accounting data only).

We measure the seven attributes on a firm- and year-specific basis, using the relevant accounting

and/or market information for rolling 10-year windows, t-9,…,t. The use of the firm as its own

benchmark mitigates concerns that differences among firms in a given industry give rise to noisy

measures of the constructs, as would be the case if we measured the attributes by reference to industry

norms. The disadvantage of the firm-specific approach relative to an industry-approach is that the former

requires a time-series of observations about each firm; the latter requires only a sufficient size cross-

section of firms in a given industry at a point in time. The time-series requirement (10 years, in our

analyses) biases our sample toward surviving firms, which are likely to be larger and more successful

than firms that do not meet this data requirement. A bias toward larger, successful firms will, if anything,

bias against finding effects in our analyses, since it likely reduces variation in the attributes we study.

10

Overall, we believe that the greater precision afforded by the use of the firm as its own benchmark offsets

the sample bias (in the cross-sectional tests) toward larger, more successful firms.

Our measure of earnings quality is based on Dechow and Dichev’s [2002] model relating current

accruals to lagged, current and future cash flows from operation:

, , 1 ,0, 1, 2, 3, ,

, , ,

j t j t j t j tj j j j

j t j t j t j t

TCA CFO CFO CFOAssets Assets Assets Assets

, 1

,j tϕ ϕ ϕ ϕ−= + + + +ν+

,

,

(1)

where = firm j’s total current accruals in year t, ;tjTCA , , , ,( )j t j t j t j tCA CL Cash STDEBT= ∆ − ∆ − ∆ + ∆ 9

,j tAssets = firm j’s average total assets in year t and t-1

,j tCFO , = cash flow from operations in year t, is calculated as net income before extraordinary items less total accruals (TA), whereTA , , , , ,j t j t j t j t j t j tCA CL Cash STDEBT DEPN= ∆ − ∆ − ∆ + ∆ −

,j tSTDEBT

,

and = firm j’s change in current assets (Compustat #4) between year t-1 and year t, = firm j’s change in current liabilities (Compustat #5) between year t-1 and year t,

= firm j’s change in cash (Compustat #1) between year t-1 and year t, ∆ = firm j’s change in debt in current liabilities (Compustat #34) between year t-1 and year t,

tjCA ,

,j t

,j t

∆

,j tCL∆

Cash∆

DEPN = firm j’s depreciation and amortization expense (Compustat #14) in year t.

For each firm-year, we estimate (1) using rolling 10-year windows. These estimations yield 10 firm- and

year-specific residuals, ,j tν , t = t-9,…,t which form the basis for the earnings quality metric,

, ˆ(j t j tQuality , )σ ν= , equal to the standard deviation of firm j’s estimated residuals. Large (small) values

of Quality correspond to poor (good) earnings quality.

Following previous research (e.g., Lev [1983]; Ali and Zarowin [1992]) we measure earnings

persistence ( ) as the slope coefficient estimate, Persistence 1, jφ , from an autoregressive model of order

one (AR1) for annual earnings:

, 0, 1, , 1 ,j t j j j tX X j tφ φ −= + +υ

(2)

9 We use the indirect (balance sheet) approach to estimate accruals rather than the direct (statement of cash flow) approach. Although the former suffers from measurement error in accruals, especially for firms with merger and acquisition activity or discontinued operations (Hribar and Collins [2002]), it allows us to calculate accruals for a larger sample of firms and over a longer period than is possible with the direct approach. In particular, the direct approach requires data from the statement of cash flows, which is not available prior to 1988, the year in which SFAS No. 95 was effective. A 10-year data requirement would, therefore, restrict our sample to firms with the necessary data in the sub-period 1999-2001.

11

For each firm-year, we estimate (2) using maximum likelihood estimation and rolling 10-year windows.

This procedure yields firm- and year-specific estimates of 1, jφ , which capture the persistence of

earnings.10 Values of 1, jφ close to one imply highly persistent earnings, while values of 1, jφ close to zero

imply highly transitory earnings. To preserve the same ordering of values across attributes, we take the

negative of 1, jφ as our measure of 1, jPersistence φ= − , so that larger (smaller) values of Persistence

correspond to less (more) persistent earnings.

Our measure of earnings predictability ( ) is also derived from the firm- and year-

specific AR1 models. Specifically, based on Lipe [1990], we use the square root of the error variance

from (2),

Predictability

2ˆ ( )jPredictability σ ν= . Large (small) values of Predictability imply less (more) predictable

earnings.

We define Smoo as the ratio of firm j’s standard deviation of net income before

extraordinary items (Compustat #18) divided by beginning total assets, to its standard deviation of cash

flows from operations divided by beginning total assets,

,j tthness

, ,( ) / ( , )j t j tSmoothness NIBE CFOj tσ σ= .11

Standard deviations are calculated over rolling 10-year windows. Larger values of Smoothness indicate

less earnings smoothness than do smaller values of Smoothness.

10 We use an AR1 model (with drift) of annual earnings, rather than a higher order specification suggested by Finger [1994] and Baginski, Lorek, Willinger and Branson [1999], because we wish to estimate firm-specific persistence measures for a broad sample of firms over rolling 10-year windows. Using higher order specifications increases the number of parameters to be estimated, and therefore, increases the length of the time-series needed for the estimation; in turn, this restricts the sample firms to those with the necessary data. For example, Finger estimates AR models of orders 2, 4 and 8 for a sample of 50 firms with at least 40 yearly observations over the period 1935-1987; Baginski et al. estimate ARIMA (2,1,0) models [among others] for 162 firms with a complete series of annual data for 1967-1990 (24 years). 11 Our measure of smoothness is the same as in Francis, LaFond, Olsson and Schipper [2003], and similar to those used by Leuz, Nanda and Wisocki [2003] and Hunt, Moyer and Shevlin [2000]. Leuz et al. examine the ratio of the standard deviation of operating income scaled by assets, to the standard deviation of cash flows from operations scaled by assets; Hunt et al. examine the ratio of the standard deviation of non-discretionary net income (equal to operating cash flows plus non-discretionary accruals) to the standard deviation of cash flows from operations.

12

Following Francis and Schipper [1999], Collins, Maydew and Weiss [1997] and Bushman, Chen,

Engel and Smith [2003], our measure of value relevance ( Relevance ) is based on the explained

variability from the following regression of returns on the level and change in earnings:

, 0, 1, , 2, , ,j t j j j t j j tRET EARN EARN j tδ δ δ ζ= + + ∆ + (3)

where firm j’s 15-month return ending 3 months after the end of fiscal year t; ,j tRET =

firm j’s earnings before extraordinary items in year t, scaled by market value at the ,j tEARN =

end of year t-1; the change in firm j’s earnings before extraordinary items in year t, scaled by market ,j tEARN∆ =

,

value at the end of year t-1. We estimate (3) for each firm over rolling 10-year windows. The negative of the adjusted R2 from (3),

for each firm- and year-specific regression, is our measure of . Large (small) values

of Relevance imply less (more) value relevant earnings.

2, , (3)j t eqRelevance R= −

Our measures of timeliness and conservatism are derived from reverse regressions, which use

earnings as the dependent variable and returns measures as independent variables:

, 0, 1, , 1, , 2, , ,j t j j j t j j t j j t j tEARN NEG RET NEG RET j tα α β β= + + + ⋅ + ς (4)

where if and 0 otherwise; and all other variables are as previously defined. , 1j tNEG = , 0j tRET <

Similar to our other attributes, equation (4) is estimated on a firm- and year-specific basis, using rolling

10-year windows. Following Ball, Kothari and Robin [2000] and Bushman et al., our measure of

timeliness is based on the explanatory power of equation (4); in particular, to preserve the same ordering

as for the other attributes, we use the negative of the adjusted R2 from (4), Tim .

Following Basu [1997], Pope and Walker [1999] and Givoly and Hayn [2000], our measure of

conservatism is the negative of the ratio of the coefficient on bad news to the coefficient on good news,

2, , (4)j t eqeliness R= −

1, 2, 1,( ) /j jConservatism jβ β β= − + . Larger values of Timeliness and Conservatism imply less timely

earnings and less conservative earnings, respectively, than do smaller values of these variables.

Information about the pooled sample distribution of the proxies for earnings attributes is reported

in Panel A, Table 3. We report this information for the Full Sample only; results are similar for the VL

13

Sample. The Quality measure has mean (median) value of 0.026 (0.019) with a standard deviation of

.023, which indicates considerable dispersion for our sample. Persistence, which captures the extent to

which an earnings innovation remains in the series, has a mean (median) value of -0.482 (-0.520), with a

standard deviation of 0.368. The 90th percentile value of 0.040 implies that at least 10% of the sample

observations are characterized by a negative relation between current and lagged earnings (recall that we

use the negative of the estimated slope coefficient to preserve the ordering of this attribute with the other

attributes we consider). Predictability has a mean (median) value of 0.876 (0.536) and a standard

deviation of 1.054, indicating both dispersion and skewness in this earnings attribute for our sample. The

last accounting-based attribute we consider, Smoothness, captures the variability of income relative to the

variability of cash flows. This ratio measure has mean (median) values of 0.640 (0.578) with a standard

deviation of 0.356.

Turning to the three market-based earnings attribute measures, we see that Relevance (the

negative of the R2 in a returns-earnings regression) has a mean (median) value of -0.423 (-0.416) with a

standard deviation of 0.243. Timeliness, the negative of the R2 in a reverse regression of earnings on

returns, is roughly comparable for our sample to value relevance, with a mean (median) value of -0.466

(-0.465) and a standard deviation of 0.243. Finally, Conservatism, the ratio of the coefficients on

negative returns to the coefficient on positive returns in a reverse regression of earnings on returns, has a

mean (median) value of 0.492 (0.000). In addition to significant skewness, this attribute measure also

exhibits substantial dispersion; the standard deviation is 24.459 and the 10th (90th) percentile is -9.141

(8.777).12

Information on the Pearson (above diagonal) and Spearman (below diagonal) pairwise

correlations among the seven attributes is reported in Panel B, Table 3 for the Full Sample (the VL

12 The large dispersion in Conservatism is driven by observations with small values of the denominator ( 1, jβ ) of

this variable, 1, 2, 1,( ) /j jnservatism jCo β β β= − + . Our tests are not affected by extreme values of Conservatism (or extreme values of other attribute proxies) because we do not use the raw values of these variables in our tests. Rather, we use the decile ranks of the variables in the cross-sectional tests (section 4), and we use factor mimicking portfolios in the time-series tests (section 5).

14

sample correlations are similar and are not reported). Turning first to the four accounting-based

attributes, we see that all of the attribute measures are positively correlated (with p-values < .0001), with

correlations exceeding 0.20 in magnitude. We also observe significant positive correlations among all

three market-based attribute measures (p-values < .0001), although only the correlation between

Relevance and Timeliness is large in economic terms, 0.67 (both Pearson and Spearman). In terms of

correlations across the two sets of attributes, the correlations between the accounting-based attributes and

Relevance and Timeliness are generally positive (p-values < .0001), with magnitudes of 0.12 or less.

Conservatism displays significantly positive Spearman correlations (of about 0.04 or less), and zero or

negative Pearson correlations, with the accounting attributes. On the whole, this evidence suggests

relatively little overlap between the accounting-based attributes and the market-based attributes.

As noted earlier, each of the earnings attributes is measured for each firm and fiscal year. To

conform these yearly observations to the quarterly VL data (used in the cross-sectional tests) and the

monthly CRSP returns (used in the time-series portfolio regression tests), we assign the value of each

yearly attribute to the months comprising that fiscal year. To ensure that these data are available to

investors, we lag the assignments by three months, to allow for the 90-day 10K filing period.13 As an

example, we estimate IBM’s value of 1, jPersistence φ= − for the fiscal year ending December 31, 1990 to

be –0.4545. We, therefore, assign Persistence = –0.4545 to all months beginning April 1991 and ending

March 1992.

4. Cross-Sectional Tests

The analyses reported in this section use the quarterly VL-based ex ante cost of equity estimates,

control variables and earnings attributes. For these tests, our control variables are risk proxies known to

influence the cost of capital: beta, firm size, and book-to-market ratio (Fama and French [1993]).14 We

13 Because equation (1), used to measure earnings quality, requires information about cash flows in year t+1, we lag the measure of earnings quality by one year to ensure that this information is known at time t. 14 Results using the CAPM as the benchmark model of expected returns (summarized in section 6) generally yield similar inferences.

15

measure firm j’s beta (Beta) in year t using the CAPM, estimated using monthly returns data over rolling

10-year windows; we require a minimum of 24 monthly returns for the CAPM estimation. Firm size

(Size) is measured as the log of firm j’s market value at the end of fiscal year t-1. The book-to-market

ratio (BM) equals the log of firm j’s book value of equity (Compustat #60) divided by its market value of

equity, both measured at the end of fiscal year t-1. Similar to the earnings attributes, we assign values of

Beta, Size and BM to each quarter with a 3-month lag to ensure the information is known to the market.

To compare coefficient estimates across attributes, we rank each attribute each year, and form

deciles. Firms in the top decile (decile 1) have the smallest values of the attribute, while firms in the

bottom decile (decile 10) have the largest values of the attribute. Given our attribute measures, this

ordering places firms with the best (worst) characterization of the attribute in the top (bottom) deciles.

We begin by examining the relationship between the implied cost of equity and earnings

attributes for each calendar quarter q, controlling for risk proxies:

, 0, 1, , 2, , 3, , 4, , ,k

j q q q j q q j q q j q q j q j qCofC Beta Size BM Attributeλ λ λ λ λ= + + + + +ς (5)

where the decile rank of firm j’s value of the k’th earnings attribute in quarter q, ,kj qAttribute =

Quality, Persistence, Predictability, Smoothness,

kRelevance, Timeliness, Conservatism

∈

We use the decile rank of each attribute, rather than its raw value, because the resulting coefficient

estimates can be interpreted as the incremental cost of equity associated with adjacent deciles; also, the

use of rank values alleviates the effects of extreme observations. To mitigate concerns about cross-

sectional dependencies in the data, we estimate (5) for each of the Q=108 calendar quarters during 1975-

2001. We report the mean of the 108 coefficient estimates, and assess statistical significance using the

time-series standard errors of the 108 estimates (Fama and MacBeth [1973]).

Table 4, panel A reports the results of estimating a base model version of (5) which excludes the

earnings attributes. As expected, Beta, Size and BM are significantly associated with cost of capital

estimates. Beta and BM are positively associated with CofC (t-statistics of 22.31 and 6.16, respectively),

16

while Size is negatively associated with (t-statistic = -7.73). Panel B reports results of estimating

(5) adding, individually, the accounting-based earnings attributes. Based on prior research which posits

that each attribute, considered individually, is viewed as desirable by investors, we expect that the

coefficient estimate on each attribute is positive,

CofC

4 0λ > , indicating higher costs of capital for firms with

the worst characterizations of the attribute (e.g., the worst earnings quality, the least persistent, the least

predictable, and the least smooth earnings).

4λ =

4λ =

Turning first to earnings quality, the results show a mean estimate of 4 0.396λ = (t-statistic =

20.02). This finding suggests that firms with the best earnings quality enjoy a 396 bp lower cost of equity

capital relative to firms with the worst earnings quality.15 The next largest effect is found for Persistence,

where the mean estimate of 4 0.216λ = (t-statistic = 10.59) suggests a spread of 216 bp between firms

with the most and least persistent earnings. The third largest cost of capital effect is associated with

Smoothness; 4 0.120λ = (t-statistic = 6.91). Finally, Predictability has the smallest effect of the

accounting-based earnings attributes we consider, with a mean value of 4 0.056λ = (t-statistic = 1.82).

Panel C reports results for each of the market-based earnings attributes. We find the largest

coefficient estimate relating CofC to Relevance, 0.114 (t-statistic = 5.39), followed by Timeliness

4 0.071λ = (t-statistic = 3.23), and Conservatism, 0.034 (t-statistic = 2.94). In terms of the cost of

equity capital, the magnitudes of these effects suggest that firms with the least value relevant (least

timely) earnings pay a premium of about 114 bp (71 bp) over firms with the most value relevant (most

timely) earnings; the firms with the least conservative earnings pay a premium of 34 bp over firms with

the most conservative earnings.

To contrast the cost of capital effects of the attributes more directly, we evaluate the incremental

contribution of each attribute, in the presence of the others, to explaining implied costs of equity:

15 The magnitude of this effect is similar to the earnings quality effect documented by Francis et al. [2002]. For the measure of Quality closest to ours, they estimate an effect of 292 bp (their estimate controls for beta, but not for firm size or the book-to-market ratio).

17

, 0, 1, , 2, , 3, , 4, , ,k k

j q q q j q q j q q j q q j q j qk

CofC Beta Size BM Attributeλ λ λ λ λ= + + + + +ς∑ (6)

Table 5 reports the results of estimating (6) for the set of accounting-based earnings attributes

(Model 1), the set of market-based earnings attributes (Model 2), and all seven earnings attributes (Model

3). The results of Model 1 show that Quality continues to dominate the other accounting-based earnings

attributes in terms of the magnitude of pricing effects: (t-statistic = 26.31), or a 417 bp

cost of capital spread between the best and worst earnings quality firms. Controlling for other

accounting-based attributes, Model 1 also shows the mean coefficient estimate on Persistence is

(t-statistic = 5.94). In the presence of the other accounting attributes, Smoothness is

insignificantly related to CofC and Predictability takes on a negative coefficient, (t-

statistic = -4.74). That is, once we take account of the Quality and Persistence earnings attributes,

favorable outcomes for the other two accounting-based attributes we consider, Predictability and

Smoothness, no longer have a reliably positive association with a reduced cost of capital.

4 0.417Qualityλ =

4 0.167Persistenceλ =

4 0.176Predictabilityλ = −

The results of Model 2 reveal that Relevance and Conservatism are distinct market-based

earnings attributes, vis-à-vis the cost of equity capital. Specifically, the mean estimate of

(t-statistic = 5.92), suggests a 118 bp spread between the worst and best value relevance

firms and the mean estimate of (t-statistic = 2.47), suggests a 28 bp spread between

the least and most conservative earnings firms. Conditioning on Relevance and Conservatism, Timeliness

is no longer associated with the cost of equity; this result is perhaps not surprising in light of the high

correlation (about 0.67, panel B of Table 3) between Relevance and Timeliness.

4 0.118Relevanceλ =

4 0.028Conservatismλ =

Results including both the accounting-based attributes and the market-based attributes are

reported in the far right columns of Table 5 (Model 3). Except for the decline in the coefficient estimate

on Relevance (from 0.118 in Model 2 to 0.071 in Model 3) and the decline in the significance of

Conservatism (coefficient estimate declines from 0.028 in Model 2 to 0.011 in Model 3 (t-statistic =

1.00), results are broadly similar to those reported for Models 1 and 2. In particular, we find the largest

18

cost of capital effects for quality ( , t-statistic = 26.41), followed by persistence

( , t-statistic = 5.95), and relevance ( , t-statistic = 3.86). Smoothness,

Timeliness and Conservatism are insignificantly associated with CofC, and Predictability retains the

negative conditional relation with the cost of capital found in Model 1.

4 0.416Qualityλ =

4 0.158Persistenceλ = 4 0.071Relevanceλ =

Tests of the incremental explanatory power of the accounting-based earnings attributes and of the

market-based earnings attributes are reported in the bottom rows of Table 5. These results show that both

sets of attributes add significantly to known risk proxies in explaining cross-sectional variation in costs of

capital. In particular, we find that the set of accounting-based attributes adds an average of 2.3 percentage

points in explanatory power (the average is calculated across the Q=108 regressions), reliably different

from zero at <.0001 level. This increment corresponds to a 16% increase over the explanatory power of

the base model (14.3% reported in panel A, Table 4). The incremental explanatory power provided by the

market-based attributes is smaller, averaging 0.5 percentage points (t-statistic = 6.01), or a 3.5% increase

over the explanatory power of the base model. In unreported tests comparing the difference in

incremental R2s provided by accounting-based attributes versus market-based attributes (that is, the mean

of 2.3 percentage points versus the mean of 0.5 percentage points), we find that accounting-based

attributes provide significantly more explanatory power.

We draw several inferences from the results of the cross-sectional tests reported in Tables 4 and

5. First, each earnings attribute we consider is individually associated with the cost of equity capital in

the predicted way. Second, accounting-based earnings attributes explain substantially more of the

variation in ex ante estimates of the cost of equity than do market-based earnings attributes. Third,

among the accounting-based earnings attributes, earnings quality is the most priced characteristic,

followed by persistence. Once we condition on quality and persistence, earnings predictability and

smoothness are valued substantially less or not at all. Fourth, among the market-based earnings

attributes, only value relevance is consistently associated with ex ante cost of capital estimates, once we

condition on all earnings attributes.

19

5. Time-Series Tests

Our time-series tests investigate the effects of earnings attributes on equity costs of capital, as

evidenced by the loadings on, and incremental explanatory of, attribute factors in firm-specific asset-

pricing regressions. We begin by calculating a factor mimicking portfolio for each of the k earnings

attributes, kmFactor , equal to the difference between the equally-weighted monthly returns of firms in the

bottom three attribute deciles (deciles 8, 9 and 10) and the equally-weighted returns of firms in the top

three deciles (deciles 1, 2 and 3).16 Following convention, we describe a factor mimicking portfolio as a

“factor.” This procedure yields a series of m=324 monthly kmFactor returns for each of the k=7 attribute

factors. We assess the importance of each attribute factor in explaining asset prices by augmenting the

standard 3-factor pricing regression with kmFactor :

, , , ,( ) k,j m F m j j Mkt m F m j m j m j m jR R a b R R s SMB h HML f Factor mζ− = + − + + + + (7)

where firm j’s return in month m; ,j mR =

,F mR = the risk free rate in month m;

,Mkt mR = the market return in month m;

mSMB = size factor in month m;

mHML = book-to-market factor in month m; k

mFactor = attribute k factor for month m. Variants of equation (7) are estimated on a firm-specific basis for all listed firms with at least 24

monthly CRSP returns. Because our tests require data only on the attribute factors (which are based on

monthly stock returns), and not on the underlying data supporting these factors (the earnings attributes

themselves), these tests are not restricted to firms in the Full Sample. Stated differently, we use data on

the earnings attributes for firms in the Full Sample to create the attribute factors, kmFactor , which can then

be correlated with the returns of any firm with returns data. The only requirement we impose in the time-

series tests is that the firm have at least 24 monthly returns observations to estimate (7); in total J=18,786 16 Our construction of the attribute factors follows Carhart’s [1997] construction of a price momentum factor. Our results are not sensitive to whether we equally-weight or value-weight securities in the attribute mimicking portfolios. Results are also not sensitive to the cutoffs used to form the factor; specifically, we draw similar inferences if we form the factor using the difference between the top and bottom 1, 2 or 4 deciles.

20

distinct firms meet this requirement (the “Traded Sample”).17 Table 1 shows that firms in the Traded

Sample comprise, on average, 99.2% of the total CRSP market capitalization.

To benchmark our results, we estimate (7) excluding all of the attribute factors. Panel A, Table 6

reports the mean coefficient estimates and adjusted R2s for the resulting 3-factor pricing regressions,

calculated across the J firm-specific regressions; t-statistics are based on the standard errors from the J

coefficient estimates. Consistent with Fama and French [1993], we find that the market risk premium

( Mkt FR R− ), the size factor (SMB) and the book-to-market factor (HML) have significant positive loadings

(t-statistics range between 21 and 156). The mean firm-specific adjusted R2 for the 3-factor regression is

14.2%.

The remaining panels in Table 6 show the results of regressions which include kmFactor

km

as an

additional independent variable; these tests allow us to assess the degree to which each attribute factor

overlaps with and adds to the known risk factors in explaining returns. Our emphasis here is on the

ability of the attribute factors to explain variation in expected returns. We, therefore, focus on the

statistical significance of the loadings on the attribute factors, not on the point estimates of those

loadings.18 Results for the accounting-based attributes (Panel B) show that, in all cases, Factor

0.525=

enters

with a significant positive loading. In particular, the mean loading on is (t-statistic

= 47.84), for (t-statistic = 41.23),

QualityFactor f

0.665f = PersistenceFactor 0.513f = for (t-statistic =

29.22), and for (t-statistic = 42.78). We also find significant positive loadings

PrFactor edictability

0.598f = SmoothnessFactor

17 The mean (median) number of observations used to estimate equation (8) is 110 (83) months. 18 Unlike the comparisons of point estimates on the decile ranks of the attributes in section 4, it is not straightforward to compare the magnitudes of the point estimates of the factor loadings because the factors themselves differ across attributes. While we note that the mean values of the factors for earnings quality and earnings persistence are noticeably larger than other attribute factors (27-29 bp per month for quality and persistence versus less than 11 bp per month for the others, not reported), we are hesitant to draw inferences from these means because, like the known risk factors, the attribute factors display considerable over-time variation. As Fama and French [1997] discuss in the context of the Mkt FR R− , SMB and HML, it is difficult to draw inferences about the magnitude of cost of capital effects when the standard error around the mean factor value is large. As a benchmark, Fama and French [1997, Table 1] report standard errors for Mkt FR R− , SMB and HML of 23 bp, 15 bp and 13 bp per month, respectively. Our attribute factors display similar variation (not reported): standard errors for the attribute factors range from 5 bp to 26 bp per month.

21

for the market-based attribute factors (Panel C). For RelevanceFactor , 0.33f 9= (t-statistic = 16.20),

compared to for (t-statistic = 9.25) and0.207f = TimelinessFactor 0.197f = for (t-statistic

= 8.54).

ConservatismFactor

,k

, , ,( ) km F mR a j Mkt m FR R m j mh HML j

kf Factorm j mζ− = + − + + +∑

QuaFactor

0.065= −

Timelinessor

lity Qualityf =

Relevancef

.532

0.341=

Smoothnessf

0.091

Qualityctor

f

PersistenceFactor

In addition to considering the effects of each attribute individually, we consider them as sets and

all together. Results based on estimations of (8) are shown in Table 7:

,j j m jR b s SMB+ (8)

Model 1 reports results of the regression which includes the four accounting-based attribute

factors. The results show the most significant loading for ( 0 , t-statistic = 25.88),

comparable, in magnitude, to the loading of 0.525 for this factor considered individually (Table 6, Panel

B) . The loadings on the other accounting-based factors are either insignificantly different from zero

( and ), or negative ( , t-statistic = -3.18). In terms of the market-

based attribute factors, Model 2 shows a significant positive value for (t-statistic =

12.50), comparable to the loading of 0.339 for this factor considered individually. The factor loading for

(t-statistic = 3.76), smaller than the loading of 0.197 for this factor considered

individually. There is no significant loading on when it is included with the other market-

based attribute factors. Finally, turning to the combined Model 3 (rightmost columns), the results show

that only and have significant positive loadings in the presence of the other

factors ( , t-statistic = 23.14, and , t-statistic = 2.45). All of the market-

based factors have negative loadings (varying between –0.053 and –0.039, with t-statistics between –1.48

and –1.85).

Persistencef

Conservatismf

f

Predicatbility

Fact

Persistencef

=

Fa

Quality 0.514= 0.057=

Table 7 also reports the increase in adjusted R2s from adding the accounting-based attributes to

the 3-factor model relative to adding the market-based attributes. The average increase in adjusted R2 is

2.3 percentage points (t-statistic = 43.96) for the accounting-based attribute factors, and 0.5 percentage

points (t-statistic = 12.67) for the market-based attribute factors. These increments correspond to

22

increases of 16% and 3%, respectively, over the explanatory power of the base 3-factor model (14.2%,

reported in panel A, Table 6). The difference in means (2.3% versus 0.5%) is significant at the .00 level

(not reported), indicating that accounting-based attribute factors add more to 3-factor pricing regressions

than do market-based attribute factors.

We draw the following inferences from the time-series tests reported in Tables 6 and 7. First, all

seven earnings attribute factors enter with the expected positive (and significant) loadings when

considered individually. When the factors are considered jointly, only the factors capturing earnings

quality and earnings persistence retain significant positive coefficients. Factors capturing relevance,

timeliness and conservatism switch sign (implying that, in the presence of the other earnings attribute

factors, firms with more relevant, more timely and more conservative earnings pay a cost of capital

premium relative to firms with less relevant, less timely and less conservative earnings). Second, across

all tests, we find that earnings quality is the most important attribute, vis-à-vis our measure of investor

resource allocation decisions, in terms of explanatory power. Third, similar to the cross-sectional tests,

we find that accounting-based earnings attributes add significantly more explanatory power to a 3-factor

model than do market-based attributes.

6. Additional Tests

In this section we report the results of three additional tests. Our first test examines the sensitivity

of the results to the choice of the benchmark model of expected returns. We repeat all tests using the

CAPM rather than the 3-factor model; a summary of the CAPM-based multivariate results is reported in

Table 8. For the accounting-based earnings attributes Quality and Persistence, relative to results based on

the 3-factor model, the CAPM cross-sectional tests show larger coefficient estimates on the attribute

proxies and the CAPM time-series tests show larger loadings on the attribute factors. Smoothness is

associated with statistically reliable negative conditional effects. Results are mixed for the market-based

earnings attributes. We note, in particular, that favorable outcomes on the attribute Relevance continue to

be associated with lower costs of equity capital, except in the time series estimation which includes all

23

earnings attributes, where this effect is reversed. We interpret these results as confirming the finding that

accounting-based earnings attributes, particularly Quality and Persistence, dominate market-based

earnings attributes in explaining cost of capital effects.

Our second test considers Penman and Zhang’s [2002] accounting-based measure of

conservatism, which is based on the presence of earnings “reserves” in the form of LIFO reserves and

unrecognized R&D and advertising assets. Their conservatism measure, C-score, equals the sum of the

LIFO reserve, estimated R&D assets and estimated advertising assets, scaled by net operating assets.19

Following Penman and Zhang, we calculate C-scores for all firms with the necessary data.20 For the Full

Sample, 26,036 firm-years (2,894 distinct firms) have data on C-scores (these firms are used to calculate

a FactorC-score for the time-series tests). For the VL sample, 14,620 firm-years (58,483 firm-quarters and

1,448 distinct firms) have data on C-scores. To provide evidence on the overlap between C-scores and

the market-based conservatism measure, we calculate pairwise correlations between the two proxies for

the sub-sample of VL firms with data on both variables. These results (not reported) show virtually no

overlap between the two proxies (correlations range between –0.0001 and 0.0123, and most are not

reliably different from zero at the 0.10 level).

We repeat the tests in Tables 4-7 replacing the market-based measure of conservatism with C-

score; a summary of the main multivariate results is reported in Table 9.21 Results of the cross-sectional

tests (panel A) show that the coefficient estimates on C-score do not differ significantly from zero. In

contrast, the market-based measure of conservatism was associated with a point estimate of 0.028 (t-

statistic = 2.47) when only market-based attributes are considered (Table 5). Turning to the time-series 19 The effect of conservatism on the quality of earnings is captured by Penman and Zhang’s Q score, Qit = Cit – [0.5Cit-1 + .5IndustrymedianCt], where Qit thus captures firm-specific current period reserves relative to last period’s reserves and the current industry median reserves. They view highly conservative earnings as being of low quality (hence undesirable), so that a zero Q value is most desirable and large positive or large negative Q values are undesirable. Results using the absolute value of Q as an accounting-based measure of conservatism show smaller and less significant cost of capital effects than the C score, so we focus only on the latter. 20 In particular, Penman and Zhang require that a firm have a positive value for at least one of the reserve items (LIFO reserve, estimated R&D assets or estimated advertising assets). 21 For brevity, we do not tabulate for this sub-sample the results of tests adding each earnings attribute, one at a time, to the 3-factor model. These results are similar to those shown in Tables 4 and 6. In particular, the coefficient estimate on the decile rank value of C-score is 0.084 (t-statistic = 3.88) and the loading on Factor C-score is 0.153 (t-statistic = 10.79).

24

tests (panel B), the loading on FactorC-score is reliably negative in Model 1, indicating that in the presence

of other accounting-based earnings attributes, firms with more conservative earnings pay a premium

relative to firms with less conservative earnings. The significant marginal negative association maintains

when all earnings attributes are considered (Model 3). Most importantly, the main findings from the prior

tables are not affected by the use of C-score: we continue to find that earnings quality dominates the other

attributes, with some evidence (from the cross-sectional tests) that earnings persistence and value

relevance are also important factors explaining variation in the cost of equity capital.

As previously noted, the earnings attributes we consider are jointly determined by management’s

reporting decisions and by innate or intrinsic factors, such as firms’ operating environments. Our third

test presents an exploratory analysis which attempts to separate these two influences on the cost of equity

capital. Specifically, we attempt to control for the intrinsic portion of earnings attributes by including the

rolling 10-year standard deviation of cash flow from operations (scaled by beginning total assets),

(CFO)σ , as a measure of the innate volatility of firms’ operating environments. We examine the

incremental effects of the earnings attributes conditional on ( )CFOσ as well as known risk factors.

Results of this investigation are reported in Table 10; panel A shows results for the cross-

sectional tests using Value Line analysts’ ex ante cost of capital estimates and panel B shows results for

the time-series tests using realized returns. Both sets of tests show that ( )CFOσ is a significant factor in

explaining the cost of equity. For example, results for the cross-sectional tests that include all seven

earnings attributes (Model 3 in Panel A) show a 367 bp difference between firms with the least and most

volatile cash flows (t-statistic = 19.05); the time-series tests also show a significant loading on

(t-statistic = 12.87). In addition, tests of incremental explanatory power show that (CFOFactorσ ) ( )CFOσ

and add between 0.2 and 1.3 percentage points in explanatory power to models which

exclude this variable (t-statistics on this increment range between 8.29 and 40.05).

(CFO)Factorσ

Conditional on ( )CFOσ as an appropriate proxy for innate volatility, the cost of capital effects of

the earnings attributes reported in Table 10 capture the marginal impacts of those attributes, associated

25

with management’s reporting decisions and incremental to effects associated with innate volatility. As

expected, the incremental effects of earnings attributes are smaller, once we control for ( )CFOσ . For

example, Model 3 (rightmost columns of panel A, Table 10) shows that firms with the worst earnings

quality pay a 223 bp premium relative to firms with the best earnings quality (t-statistic = 12.46),

conditional on known risk factors and ( )CFOσ . This effect is about half as large as that documented in

Table 5 (416 bp). Inspection of the results in Table 10 shows that the inclusion of (CF )Oσ has the

largest effects on the coefficient estimates and factor loadings for earnings quality (where they decrease)

and smoothness (where they increase).22 Conditioning on ( )CFOσ has little effect on the magnitude or

significance of the coefficient estimates for the market-based earnings attributes. On the whole, these

results suggest that while innate characteristics of firms’ operating environments explain a significant

portion of the variation in the cost of equity capital, a material amount of variation remains to be

explained by other factors, the most important of which is the earnings quality attribute.

σ) (CFO)σ

( )CFO

7. Summary and Conclusions

We draw on prior research to identify earnings attributes that are posited as desirable or

advantageous, with the implication that the financial reporting function (including standard setters,

preparers, auditors and enforcement authorities) should strive to report earnings with these attributes. The

attributes we consider – quality, persistence, predictability, smoothness, value relevance, timeliness and

conservatism – have been used individually or sometimes in pairs in prior studies to capture beneficial

features of earnings. However, the benefits of the attributes are typically specific to the setting

considered, so it is not possible to assess whether the earnings attributes are statistically and economically

)22 The pairwise correlation between the decile ranks (and raw values) of (CFO and Quality is about 0.60 (p-value

< .0001), and about –0.20 (p-value < .0001) between (CFOσ and Smoothness. Because is the

denominator of Smoothness = ( )NIBEσσ , the interpretation of results for the Smoothness variable

controlling for (CFO)σ differs from results which do not control for ( )CFOσ (reported in Tables 4-7). Specifically, conditional on ( )CFOσ , Smoothness in Table 10 captures the volatility of earnings.

26

distinct, and whether one or a few of the attributes dominate the others. We calibrate these earnings

attributes against a single summary outcome indicator of investors’ capital allocation decisions (the cost

of equity capital) to learn which attributes are viewed by investors as conferring the greatest capital

market advantage, as measured by a decreased cost of equity capital.

We use two complementary estimation approaches and samples. The first approach applies

cross-sectional estimation to a sample of Value Line firms. For this approach, our cost of capital

measures are ex ante estimates derived from Value Line analysts’ target share prices and their forecasts of

dividends and dividend growth rates. The second approach uses firm-specific time-series estimations of

3-factor asset pricing regressions augmented by earnings attributes. Because we use attribute mimicking

portfolios and not the attribute measures themselves, the time-series sample includes all firms with

sufficient returns to estimate the firm-specific asset pricing regressions.

Consistent with expectations based on previous research, we find a statistically reliable

association, in both cross-sectional and time-series tests, between each attribute considered individually

and measures of the cost of equity capital. When we estimate conditional associations that include all

seven earnings attributes, we find that the accounting-based earnings attributes (quality, persistence,

predictability and smoothness) dominate the market-based attributes (value relevance, timeliness and

conservatism). In cross-sectional tests, both earnings quality and earnings persistence have strong

conditional effects on the cost of equity capital, as does the market-based attribute value relevance. In

time-series tests, only the two accounting-based attributes, quality and persistence, have significant

conditional effects. We also find conditional negative effects for predictability in both the cross-sectional

and time-series specifications, and for all three market-based attributes in the time-series specification.

These findings are robust to several sensitivity checks including the choice of benchmark model

of expected returns (3-factor versus CAPM) and the use of an accounting-based versus a market-based

measure of conservatism. When we control for the innate volatility of the firm’s operating environment,

as proxied by the standard deviation of cash flows, we find that earnings quality continues to have a

statistically reliable association with the cost of equity capital, although the effect is about half that

27

documented in our other tests. Once we control for innate volatility, results for persistence and value

relevance are sensitive to the specification (cross-sectional versus time series estimation). The weight of

the evidence suggests that, among the seven attributes we consider, earnings quality is the dominant

attribute in terms of cost of capital effects, followed by earnings persistence and value relevance.

Conditional on our use of the equity cost of capital to summarize investors’ capital allocation

decisions, and on the specifics of our earnings attribute measures, we draw two inferences from these

results. First, with respect to research that uses earnings attributes to compare earnings numbers (or more

generally the reporting systems which produce the numbers), our results suggest that a focus on

accounting-based attributes produces sharper results than a focus on market-based attributes. The two

kinds of measures differ in their (implicit) assumptions about the function of earnings. That is, measures

of accounting-based earnings attributes take cash or earnings itself as the reference construct, so the

implicit assumption is that earnings are supposed to spread cash receipts and disbursements over reporting

periods. In contrast, measures of market-based attributes take returns or prices as the reference construct,

and assume that earnings should reflect the information in stock price changes. Second, with respect to

the financial reporting function, our results suggest that the user or investor perspective on earnings is

better captured by the accounting-based attributes earnings quality and persistence, possibly coupled with

the market-based attribute value relevance, than by the accounting-based attributes predictability and

smoothness and the market-based attributes conservatism and timeliness.

28

Table 1

Descriptive Information on Full Sample, VL Sample and Traded Sample Full Sample Value Line Sample Traded Sample

Year # firms % Market Cap. # firms % Market Cap. # firms % Market Cap. 1975 678 51.9% 524 49.5% 4,760 99.0% 1976 809 52.7% 604 50.1% 4,988 99.8% 1977 923 51.9% 675 49.4% 5,138 99.9% 1978 995 50.2% 710 47.5% 5,082 99.9% 1979 1,047 51.3% 739 47.8% 5,018 99.9% 1980 1,088 50.2% 772 46.8% 5,153 99.7% 1981 1,107 53.6% 778 50.2% 5,498 99.7% 1982 1,113 52.4% 769 49.3% 5,654 99.7% 1983 1,461 58.5% 848 53.9% 6,168 99.5% 1984 1,708 63.3% 1,022 58.6% 6,477 99.7% 1985 1,677 61.0% 990 56.4% 6,581 99.1% 1986 1,636 59.8% 926 53.5% 6,983 99.6% 1987 1,544 59.9% 897 52.6% 7,329 99.6% 1988 1,520 59.2% 873 51.9% 7,377 99.7% 1989 1,523 56.8% 834 48.7% 7,194 99.7% 1990 1,490 57.4% 781 48.1% 7,081 99.8% 1991 1,514 54.4% 775 46.5% 7,106 99.6% 1992 1,577 51.5% 779 45.0% 7,397 99.5% 1993 1,614 47.7% 775 41.1% 7,777 98.8% 1994 1,735 49.0% 814 42.1% 8,370 99.5% 1995 1,828 47.2% 801 40.1% 8,681 99.3% 1996 1,862 48.3% 782 40.2% 9,237 99.1% 1997 1,965 48.5% 812 42.2% 9,453 99.4% 1998 1,997 49.3% 801 42.7% 9,349 99.4% 1999 1,904 46.8% 770 40.0% 8,923 97.7% 2000 1,809 47.9% 777 41.7% 7,996 95.8% 2001 1,602 46.3% 706 43.1% 7,092 95.4%

Mean 1,471 52.9% 790 47.4% 6,958 99.2%

Distinct 3,917 1,865 18,786 Sample descriptions and variable definitions: The Full Sample contains firms with data on all seven earnings attributes in a given year t, t=1975-2001. The Value Line (VL) Sample contains the sub-set of firms in the Full Sample with data on ex ante costs of capital from Value Line. The Traded Sample consists of all securities with at least 24 monthly CRSP returns over 1975-2001. % Market Cap is the ratio of the market value of the indicated sample to the market value of the total CRSP population with share price and shares traded data available in December of the indicated year.

29

Table 2

Descriptive Statistics on Ex Ante Cost of Capital Estimates (in Percent)a

Year # firms mean std. dev. 10% 25% median 75% 90% 1975 524 33.19 12.12 18.56 24.67 32.25 40.86 50.02 1976 604 29.48 10.09 17.17 22.49 28.36 35.93 42.71 1977 675 29.08 8.61 18.24 23.46 28.29 34.51 40.43 1978 710 29.18 7.61 20.03 24.17 29.06 33.48 38.60 1979 739 31.97 9.01 20.48 26.20 32.23 37.87 43.56 1980 772 32.46 10.95 18.00 25.57 33.04 38.75 46.12 1981 778 29.05 7.24 19.97 24.14 28.68 33.56 38.41 1982 769 31.20 8.29 21.34 25.83 30.97 36.25 41.55 1983 848 19.97 5.99 12.24 16.16 19.92 23.91 27.81 1984 1,022 22.61 6.82 14.03 18.09 22.44 26.91 31.29 1985 990 19.88 9.86 10.74 14.84 19.35 23.26 27.96 1986 926 15.19 6.73 7.69 11.05 14.58 18.41 22.44 1987 897 14.18 5.78 7.56 10.65 13.59 16.75 21.14 1988 873 17.98 6.72 10.79 13.90 16.92 21.24 26.19 1989 834 16.47 6.55 9.88 12.53 15.67 19.39 24.17 1990 781 19.81 8.07 10.64 14.00 18.79 24.01 30.72 1991 775 18.28 7.96 9.38 12.66 17.12 22.42 27.87 1992 779 16.88 7.57 7.82 11.50 16.21 21.46 26.35 1993 775 14.38 6.64 6.13 9.62 14.13 18.35 22.66 1994 814 15.20 6.10 8.04 10.81 14.62 18.61 22.66 1995 801 14.78 5.85 7.87 10.67 14.25 17.83 22.11 1996 782 13.87 6.37 6.58 9.31 12.88 17.20 22.39 1997 812 12.41 5.74 6.13 8.52 11.47 15.00 20.23 1998 801 13.38 7.35 5.07 8.21 12.17 17.15 22.89 1999 770 15.99 8.28 6.33 10.12 15.41 20.56 26.43 2000 777 19.01 9.23 8.15 12.52 18.33 24.37 31.04 2001 706 16.63 8.00 8.07 11.18 15.28 20.50 26.46

Mean 790 20.83 7.76 11.74 15.66 20.22 25.13 30.53

a We report summary statistics on the yearly distribution of the ex ante cost of capital estimates, CofC, for the Value Line sample; see Table 1 for a description of this sample. CofC estimates are the values of CofC that solve the

following equality:

4 4

4

(1 ) (1 )

(1 )

CofC gDIVCofC gTPCofC

P P

+ − + − + = + , where TP = VL 4-year out target price; DIV =

VL forecast of next period dividends, and g = VL forecast of growth rate of dividends, and P = stock price nine days prior to the date of the VL report.

30

Table 3

Summary Information on Accounting-Based and Market-Based Earnings Attributes Panel A: Summary information on the distributions of earnings attributesa Attribute mean std. dev. 10% 25% median 75% 90% Quality 0.026 0.023 0.006 0.011 0.019 0.033 0.055 Persistence -0.482 0.368 -0.940 -0.809 -0.520 -0.223 0.040 Predictability 0.876 1.054 0.184 0.300 0.536 0.989 1.847 Smoothness 0.640 0.356 0.234 0.363 0.578 0.854 1.118 Relevance -0.423 0.243 -0.762 -0.615 -0.416 -0.222 -0.093 Timeliness -0.466 0.243 -0.801 -0.661 -0.465 -0.271 -0.132 Conservatism 0.492 24.459 -9.141 -1.702 0.000 2.073 8.777 Panel B: Pairwise correlations between attributesb Quality Persistence Predictability Smoothness Relevance Timeliness ConservatismQuality 1.0000 0.2899 0.2026 0.4797 0.0528 0.0171 -0.0199 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 Persistence 0.3586 1.0000 0.2338 0.2238 0.1227 0.0822 0.0025 <.0001 <.0001 <.0001 <.0001 <.0001 0.0957 Predictability 0.2691 0.3793 1.0000 0.2722 0.0904 0.0655 -0.0184 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 Smoothness 0.4610 0.2616 0.3054 1.0000 0.0959 0.0738 -0.0167 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 Relevance 0.0378 0.1223 0.1160 0.0968 1.0000 0.6770 0.0204 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 Timeliness -0.0022 0.0793 0.0844 0.0740 0.6748 1.0000 0.0081 0.1495 <.0001 <.0001 <.0001 <.0001 <.0001 Conservatism 0.0167 0.0411 0.0250 0.0146 0.0885 0.0312 1.0000 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 Sample description and variable definitions: The sample consists of all firm-years in the Full Sample; see Table 1 for a description of this sample. All variables are measured each year for each firm, using rolling 10-year windows. Quality = the standard deviation of firm j’s residuals from a regression of current accruals on lagged, current and future cash flows from operations; Persistence = the negative of firm j’s slope coefficient from an AR1 model of annual earnings; Predictability = the square root of the error variance from firm j’s AR1 model; Smoothness = the ratio of firm j’s standard deviation of earnings before extraordinary items (scaled by assets) to the standard deviation of cash flows from operations (scaled by assets); Relevance = the negative of the adjusted R2 from a regression of 15-month returns on the level and change in annual earnings (before extraordinary items); Timeliness = the negative of the adjusted R2 from a reverse regression of annual earnings (before extraordinary items) on variables capturing positive and negative 15-month returns; Conservatism = the negative of the ratio of the coefficient on bad news (negative returns) to good news (positive returns) in the reverse regression. a We report the mean values of each statistics calculated across all firm-years.

b Pearson (Spearman) correlations are reported above (below) the diagonal. Significance levels are shown in italics.

31

Table 4

Results of Quarterly Cross-Sectional Regressions of Ex Ante Costs of Equity on Risk Proxies and Earnings Attributes

Panel A: Base model regression resultsa Indep. Var. coef. t-stat. Beta 5.546 22.31 Size -0.610 -7.73 BM 1.211 6.16

Adj. R2 0.143 Panel B: Base model, plus accounting-based attributesa Decile rank value of accounting-based attribute Quality Persistence Predictability Smoothness Indep. Var. coef. t-stat. coef. t-stat. coef. t-stat. coef. t-stat. Beta 4.450 19.82 5.473 21.60 5.403 22.85 5.423 22.15 Size -0.352 -4.82 -0.528 -6.42 -0.611 -7.40 -0.611 -7.86 BM 1.457 7.52 0.834 4.20 1.149 4.77 1.168 5.90 Attribute 0.396 20.02 0.216 10.59 0.056 1.82 0.120 6.91

Adj. R2 0.158 0.148 0.151 0.147 Panel C: Base model, plus market-based attributesa Decile rank value of market-based attribute Relevance Timeliness Conservatism Indep. Var. coef. t-stat. coef. t-stat. coef. t-stat. Beta 5.543 22.64 5.528 22.32 5.546 22.39 Size -0.623 -8.08 -0.615 -8.02 -0.609 -7.69 BM 1.194 6.17 1.175 6.08 1.208 6.14 Attribute 0.114 5.39 0.071 3.23 0.034 2.94

Adj. R2 0.147 0.148 0.143 Sample description and variable definitions: The sample consists of all firm-quarters included in the VL sample; see Table 1 for a description of this sample. CofC = ex ante cost of equity estimate; Beta = firm j’s CAPM beta; Size = log of firm j’s market value; BM = log of firm j’s book-to-market ratio. Attribute = the decile rank of the noted attribute, see Table 3 for definitions of the attributes. a Each quarter, q=1,…,108, we estimate the cross-sectional relation between CofC, known risk proxies (Beta, Size and BM), and the decile ranks of the earnings attributes considered separately. We report the mean of the coefficient estimates, across the Q=108 quarters; t-statistics are based on the standard errors of the time-series of Q estimates. Panel A shows results excluding all attributes; Panel B shows results for the accounting-based attributes considered individually; and Panel C shows results for the market-based attributes considered individually.

32

Table 5

Results of Cross-Sectional Regressions of Ex ante Costs of Equity Risk Proxies and Sets of Accounting-Based and Market-Based Earnings Attributes

Model 1 Model 2 Model 3 Indep. Var.a coef. t-stat. coef. t-stat. coef. t-stat. Beta 4.403 19.57 5.501 22.46 4.382 19.44 Size -0.186 -2.16 -0.620 -8.05 -0.197 -2.38 BM 1.553 6.68 1.159 5.98 1.519 6.62 Quality 0.417 26.31 -- -- 0.416 26.41 Persistence 0.167 5.94 -- -- 0.158 5.95 Predictability -0.176 -4.74 -- -- -0.167 -4.63 Smoothness 0.003 0.19 -- -- -0.002 -0.16 Relevance -- -- 0.118 5.92 0.071 3.86 Timeliness -- -- -0.012 -0.54 -0.008 -0.38 Conservatism -- -- 0.028 2.47 0.011 1.00

Adj. R2 0.168 0.150 0.173 Tests of incremental explanatory powerb Inc R2 t-stat. p-value Model 3 versus Model 2 0.023 16.91 <.0001 Model 3 versus Model 1 0.005 6.01 <.0001 Sample description and variable definitions: See Table 4. a Each quarter, q=1,…,108, we estimate the cross-sectional relation between CofC, known risk factors (Beta, Size and BM), and the decile ranks of the earnings attributes considered jointly. We report the mean of the coefficient estimates, across the Q=108 quarters; t-statistics are based on the standard errors of the time-series of Q estimates. Model 1 shows results for the set of accounting-based attributes; Model 2 shows results for the set of market-based attributes; and Model 3 shows results for all attributes. b Each quarter Q=108, we assess the incremental explanatory power of the accounting-based attributes (Model 3 versus Model 2) and of the market-based attributes (Model 3 versus 1). We report the mean incremental explanatory power calculated across the Q estimates, along with t-statistics of whether that mean difference is reliably different from zero.

33

Table 6

Results of Time-Series Tests of the Market Pricing of Known Risk Factors and Accounting-Based and Market-Based Earnings Attribute Factors, Considered Individually

Panel A: Base model, time-series regression resultsa Indep. Var. coef. t-stat.

Mkt FR R− 0.917 155.58 SMB 0.948 105.56 HML 0.225 21.39

Adj. R2 0.142 Panel B: Based model, plus factors for accounting-based attributesa Factor based on accounting-based attribute Quality Persistence Predictability Smoothness Indep. Var. coef. t-stat. coef. t-stat. coef. t-stat. coef. t-stat.

Mkt FR R− 0.847 138.98 0.923 155.94 0.824 119.77 0.852 137.34 SMB 0.538 47.87 0.674 62.10 0.893 96.91 0.742 77.69 HML 0.365 34.98 0.141 12.71 0.149 13.24 0.328 31.55 Factor 0.525 47.84 0.665 41.23 0.513 29.22 0.598 42.78

Adj. R2 0.158 0.150 0.150 0.154 Panel C: Base model, plus factors for market-based attributesb Factor based on market-based attribute Relevance Timeliness Conservatism Indep. Var. coef. t-stat. coef. t-stat. coef. t-stat.

Mkt FR R− 0.909 151.28 0.890 141.40 0.910 151.76 SMB 0.964 103.68 0.937 96.90 0.940 104.32 HML 0.261 24.37 0.253 23.56 0.220 20.66 Factor 0.339 16.20 0.207 9.25 0.197 8.54

Adj. R2 0.145 0.146 0.144 Sample description and variable definitions: The sample consists of J=18,786 firms in the Traded Firms sample; see Table 1 for a description of this sample. Mkt FR R− =market risk premium; SMB = return to the size mimicking portfolio; HML = return to the book-to-market mimicking portfolio; kFactor = return to the k’th attribute mimicking portfolio, k = Quality, Persistence, Predictability, Smoothness, Relevance, Timeliness, Conservatism. a We estimate firm-specific regressions of firm j’s monthly excess return on the known risk factors ( Mkt FR R− , SMB and HML) and the attribute factors, kFactor . We report the mean coefficient estimates across the J=18,786 firms in the sample; t-statistics are based on the standard errors across the J coefficient estimates. Panel A shows results excluding all attribute factors; Panel B shows results for the accounting-based attribute factors considered individually; and Panel C shows results for the market-based attribute factors considered individually.

34

Table 7

Results of Time-Series Tests of the Market Pricing of Known Risk Factors and Accounting-Based and Market-Based Earnings Attribute Factors, Considered Jointly

Model 1 Model 2 Model 3 Indep. Var.a coef. t-stat. coef. t-stat. coef. t-stat.

Mkt FR R− 0.837 118.82 0.877 134.50 0.837 114.40 SMB 0.508 41.52 0.923 93.44 0.493 36.56 HML 0.343 26.11 0.239 21.70 0.317 22.03 Quality 0.532 25.88 -- -- 0.514 23.14 Persistence 0.033 1.45 -- -- 0.057 2.45 Predictability -0.065 -3.18 -- -- -0.030 -1.39 Smoothness -0.020 -0.86 -- -- 0.003 0.12 Relevance -- -- 0.341 12.50 -0.053 -1.71 Timeliness -- -- -0.010 -0.36 -0.054 -1.85 Conservatism -- -- 0.091 3.76 -0.039 -1.48

Adj. R2 0.169 0.151 0.174 Tests of incremental explanatory powerb Inc R2 t-stat. p-value Model 3 versus Model 2 0.023 43.96 <.0001 Model 3 versus Model 1 0.005 12.67 <.0001 Sample description and variable definitions: See Table 6. a For each of the J=18,786 firms in the Traded Sample, we estimate a regression of monthly excess returns on the known risk factors ( Mkt FR R− , SMB and HML), and the attribute factors, kFactor . We report the mean of the coefficient estimates, across the J estimates; t-statistics are based on the standard errors of the J estimates. Model 1 shows results for the set of accounting-based attribute factors; Model 2 shows results for the set of market-based attribute factors; and Model 3 shows results for all attribute factors. b For each firm (J=18,786), we assess the incremental explanatory power of the accounting-based attribute factors (Model 3 versus Model 2) and of the market-based attribute factors (Model 3 versus 1). We report the mean incremental explanatory power across the J estimates, along with t-statistics of whether that mean difference is reliably different from zero.

35

Table 8

Multivariate Results of Cross-Sectional and Time-Series Tests of the Cost of Capital Effects Of Earnings Attributes, Using the CAPM as the Benchmark Model of Expected Returnsa

Panel A: Mean values of coefficient estimates from quarterly cross-sectional regressions Model 1 Model 2 Model 3 Indep. Var. coef. t-stat. coef. t-stat. coef. t-stat. Beta 4.232 15.21 5.935 18.01 4.237 15.38 Quality 0.426 15.94 -- -- 0.424 16.60 Persistence 0.289 7.75 -- -- 0.281 7.82 Predictability -0.087 -3.13 -- -- -0.076 -2.70 Smoothness -0.044 -2.66 -- -- -0.046 -2.85 Relevance -- -- 0.097 4.35 0.047 2.26 Timeliness -- -- -0.011 -0.41 -0.001 -0.05 Conservatism -- -- 0.049 4.30 0.026 2.29

Adj. R2 0.127 0.095 0.134 Tests of incremental explanatory power Inc R2 t-stat. p-value Model 3 versus Model 2 0.039 14.32 <.0001 Model 3 versus Model 1 0.008 8.16 <.0001 Panel B: Factor loadings from calendar-time-regressions including attribute mimicking factors Model 1 Model 2 Model 3 Indep. Var. coef. t-stat. coef. t-stat. coef. t-stat.

Mkt FR R− 0.769 113.07 0.910 141.74 0.785 109.59 Quality 0.668 40.06 -- -- 0.614 32.68 Persistence 0.375 18.31 -- -- 0.385 17.77 Predictability 0.024 1.22 -- -- 0.039 1.92 Smoothness -0.228 -10.02 -- -- -0.094 -3.70 Relevance -- -- 0.635 22.70 -0.268 -8.93 Timeliness -- -- -0.365 -12.76 -0.200 -7.04 Conservatism -- -- -0.001 -0.06 -0.110 -4.25

Adj. R2 0.152 0.113 0.159 Tests of incremental explanatory power Inc R2 t-stat. p-value Model 3 versus Model 2 0.046 74.25 <.0001 Model 3 versus Model 1 0.007 17.69 <.0001 Sample description and variable definitions: See Tables 4 and 6. a The tests reported in panel A (panel B) are identical to those in Table 5 (Table 7) except we use the CAPM as the benchmark model of expected returns.

36

Table 9

Multivariate Results of Cross-Sectional and Time-Series Tests of the Cost of Capital Effects Of Earnings Attributes, Using Accounting-Based Measure of Conservatism

Panel A: Mean values of coefficient estimates from quarterly cross-sectional regressions Model 1 Model 2 Model 3 Indep. Var. coef. t-stat. coef. t-stat. coef. t-stat. Beta 3.503 14.05 4.173 16.03 3.502 14.10 Size -0.137 -1.58 -0.569 -7.46 -0.157 -1.90 BM 1.768 7.51 1.457 7.51 1.720 7.44 Quality 0.316 17.08 -- -- 0.313 16.86 Persistence 0.278 11.15 -- -- 0.268 11.48 Predictability -0.272 -7.54 -- -- -0.261 -7.33 Smoothness 0.035 2.08 -- -- 0.033 1.99 Relevance -- -- 0.097 4.35 0.057 2.60 Timeliness -- -- 0.000 -0.01 -0.001 -0.02 C-Score 0.010 0.46 -- -- 0.014 0.67

Adj. R2 0.142 0.122 0.147 Tests of incremental explanatory power Inc R2 t-stat. p-value Model 3 versus Model 2 0.025 16.17 <.0001 Model 3 versus Model 1 0.005 5.78 <.0001 Panel A: Factor loadings from calendar-time-regressions including attribute mimicking factors Model 1 Model 2 Model 3 Indep. Var. coef. t-stat. coef. t-stat. coef. t-stat.

Mkt FR R− 0.859 126.18 0.887 140.46 0.857 121.87 SMB 0.519 41.92 0.914 96.06 0.509 38.72 HML 0.217 16.32 0.226 21.11 0.211 14.87 Quality 0.654 30.36 -- -- 0.643 28.32 Persistence 0.005 0.22 -- -- 0.007 0.28 Predictability 0.073 4.29 -- -- 0.081 4.59 Smoothness -0.129 -5.21 -- -- -0.136 -5.35 Relevance -- -- 0.239 10.49 -0.014 -0.57 Timeliness -- -- -0.095 -3.87 -0.051 -1.98 C-Score -0.248 -13.99 -- -- -0.231 -12.55

Adj. R2 0.169 0.147 0.172 Tests of incremental explanatory power Inc R2 t-stat. p-value Model 3 versus Model 2 0.025 44.57 <.0001 Model 3 versus Model 1 0.003 10.43 <.0001 Variable definitions: See Tables 4 and 6. a The tests reported in panel A (panel B) are identical to those in Table 5 (Table 7) except we replace the market-based measure of conservatism with the accounting-based measure, C-score. Also, because data on C-score are not available for all VL firms, the sample used in the cross-sectional tests (panel A) differs from that reported in in Table 5.

37

Table 10

Multivariate Results of Cross-Sectional and Time-Series Tests of the Cost of Capital Effects Of Earnings Attributes, Controlling for the Volatility of Cash Flows

Panel A: Mean values of coefficient estimates from quarterly cross-sectional regressionsa Model 1 Model 2 Model 3 Indep. Var. coef. t-stat. coef. t-stat. coef. t-stat. Beta 3.846 17.38 4.498 19.10 3.824 17.26 Size -0.047 -0.56 -0.372 -5.22 -0.059 -0.73 BM 1.829 7.83 1.465 7.60 1.805 7.80 Quality 0.225 12.52 -- -- 0.223 12.46 Persistence 0.169 6.33 -- -- 0.159 6.30 Predictability -0.259 -7.34 -- -- -0.253 -7.63 Smoothness 0.187 10.82 -- -- 0.184 10.95 Relevance -- -- 0.113 5.74 0.076 4.16 Timeliness -- -- -0.004 -0.19 0.001 0.06 Conservatism -- -- 0.022 1.95 0.010 0.90

( )CFOσ 0.363 18.51 0.354 19.75 0.367 19.05

Adj. R2 0.176 0.163 0.180 Inc R2 t-stat. Inc R2 t-stat. Inc R2 t-stat. Adding ( )CFOσ 0.008 8.29 0.013 9.59 0.007 8.38 Panel B: Factor loadings from calendar-time-regressions including attribute mimicking factorsb Model 1 Model 2 Model 3 Indep. Var. coef. t-stat. coef. t-stat. coef. t-stat.

Mkt FR R− 0.836 117.25 0.856 131.46 0.834 112.05 SMB 0.484 37.31 0.489 37.06 0.463 32.92 HML 0.332 24.72 0.342 31.13 0.289 19.26 Quality 0.210 6.26 -- -- 0.131 3.70 Persistence 0.010 0.43 -- -- 0.018 0.73 Predictability -0.050 -2.46 -- -- 0.015 0.68 Smoothness 0.018 0.72 -- -- 0.049 1.76 Relevance -- -- -0.130 -4.41 -0.045 -1.42 Timeliness -- -- 0.064 2.31 -0.112 -3.81 Conservatism -- -- -0.044 -1.83 0.012 0.44

( )CFOσ 0.342 11.57 0.596 42.60 0.406 12.87

Adj. R2 0.171 0.164 0.176 Inc R2 t-stat. Inc R2 t-stat. Inc R2 t-stat. Adding ( )CFOσ 0.002 11.35 0.013 40.05 0.002 10.35

38

Sample description and variable definitions: See Tables 4 and 6; also (CFO)σ = standard deviation of the firm’s cash flows from operations, scaled by total assets, measured over rolling 10-year windows. a Each quarter, q=1,…,108, we estimate the cross-sectional relation between CofC, known risk factors (Beta, Size and BM), the decile rank of (CFO)σ , and the decile ranks of the earnings attributes considered jointly. We report the mean of the coefficient estimates, across the Q=108 quarters; t-statistics are based on the standard errors of the time-series of Q estimates. Model 1 shows results for the set of accounting-based attributes; Model 2 shows results for the set of market-based attributes; and Model 3 shows results for all attributes. We also report the mean (across the Q estimates) incremental explanatory power of (CFO)σ for each model, along with t-statistics of whether that mean difference is reliably different from zero. b For each of the J=18,786 firms in the Traded Sample, we estimate a regression of monthly excess returns on the known risk factors ( Mkt FR R− , SMB and HML), a cash flow volatility factor, , and the attribute factors, (CFOFactorσ )

kFactor . We report the mean of the coefficient estimates, across the J estimates; t-statistics are based on the standard errors of the J estimates. Model 1 shows results for the set of accounting-based attribute factors; Model 2 shows results for the set of market-based attribute factors; and Model 3 shows results for all attribute factors. We also report the mean (across the J estimates) incremental explanatory power of for each model, along with t-statistics of whether that mean difference is reliably different from zero.

(CFFactorσ )O

39

References Ali, A. and P. Zarowin. 1992. The role of earnings levels in annual earnings-returns studies. Journal of Accounting Research 30(2): 286-296. American Institute of Certified Public Accountants (AICPA). 1994. Comprehensive Report of the Special Committee on Financial Reporting. New York: American Institute of Certified Public Accountants. Association for Investment Management and Research (AIMR). 1993. Financial Reporting in the 1990s and Beyond. Charlottesville, VA: Association for Investment Management and Research. Baginski, S., K. Lorek, G.L. Willinger and B. Branson. 1999. The relationship between economic characteristics and alternative annual earnings persistence measures. The Accounting Review 74, 1: 105-120. Ball, R., S.P. Kothari and A. Robin. 2000. The effect of international institutional factors on properties of accounting earnings. Journal of Accounting and Economics 29, 1: 1-51. Barth, M., W. Beaver and W. Landsman. 2001. The relevance of the value relevance literature for financial accounting standard setting: another view. Journal of Accounting and Economics 31 (1-3): 77-104. Basu, S. 1997. The conservatism principle and the asymmetric timeliness of earnings. Journal of Accounting and Economics 24, 1: 3-37. Bhattacharya, U., H. Daouk and M. Welker. 2003. The world pricing of earnings opacity. The Accounting Review, forthcoming. Botosan, C. 1997. Disclosure level and the cost of equity capital. The Accounting Review 72: 323-349. Botosan, C. and M. Plumlee. 2002. A re-examination of disclosure level and the expected cost of equity capital. Journal of Accounting Research 40, 1: 21-40. Brav, A. and J. B. Heaton. 2002. Competing theories of financial anomalies. The Review of Financial Studies 15,2 : 575-606. Brav, A., R. Lehavy and R. Michaely. 2002. Expected returns and asset pricing. Working paper, Duke University, University of Michigan, Cornell University. Bushman, R., Q. Chen, E. Engel and A. Smith. 2002. Financial accounting information, organizational complexity and corporate governance systems. Working paper, University of North Carolina, Duke University and University of Chicago. Carhart, M. 1997. On persistence in mutual fund performance. Journal of Finance 52, 1: 57-82. Chaney, P. and C. Lewis. 1995. Earnings management and firm valuation under asymmetric information. Journal of Corporate Finance: Contracting, governance and organization 1, 319-345. Collins, D. and S.P. Kothari. 1989. An analysis of inter-temporal and cross-sectional determinants of earnings response coefficients. Journal of Accounting and Economics 11: 143-181.

40

Collins, D., E. Maydew and I. Weiss. 1997. Changes in the value-relevance of earnings and book values over the past forty years. Journal of Accounting and Economics 24: 39-67. Dechow, P. and I. Dichev. 2002. The quality of accruals and earnings: The role of accrual estimation errors. The Accounting Review 77, supplement: 35-59. Easley, D., S. Hvidkjær and M. O’Hara. 2002. Is information risk a determinant of asset returns? Journal of Finance 57:5 (October): 2185-2221. Easley, D. and M. O’Hara. 2001. Information and the cost of capital. Cornell University working paper. Easton, P. and M. Zmijewski. 1989. Cross-sectional variation in the stock market response to accounting earnings announcements. Journal of Accounting and Economics 11, 117-141. Elton, E. 1999. Expected return, realized return, and asset pricing tests. Journal of Finance 54, 4: 1199-1220. Fama, E. and K. French. 1993. Common risk factors in the returns on stocks and bonds. Journal of Financial Economics 33: 3-56. Fama, E. and K. French. 1997. Industry costs of equity. Journal of Financial Economics 43: 153-193. Fama, E. and J. MacBeth. 1973. Risk, return and equilibrium: empirical tests. Journal of Political Economy 81: 607-636. Finger, C. 1994. The ability of earnings to predict future earnings and cash flow. Journal of Accounting Research 32, 2: 210-223. Francis, J. and K. Schipper. 1999. Have financial statements lost their relevance? Journal of Accounting Research 37: 319-352. Francis, J., R. LaFond, P. Olsson and K. Schipper. 2002. The market pricing of earnings quality. Working paper, Duke University, University of Wisconsin and the FASB. Francis, J., R. LaFond, P. Olsson and K. Schipper. 2003. Earnings quality and the pricing effects of earnings patterns. Working paper, Duke University, University of Wisconsin and the FASB. Gebhardt, W., C. Lee and B. Swaminathan. 2001. Toward an implied cost of capital. Journal of Accounting Research 39, 1: 135-176. Givoly, D. and C. Hayn. 2000. The changing time-series properties of earnings, cash flows and accruals: Has financial reporting become more conservative? Journal of Accounting and Economics 29, 3: 287-320. Harris, T., E. Huh and P. Fairfield. 2000. Gauging profitability on the road to valuation. Strategy Report, Global Valuation and Accounting, Morgan Stanley Dean Witter. Hribar, P. and D. Collins. 2002. Errors in estimating accruals: Implications for empirical research. Journal of Accounting Research 40, 1: 105-134.

41

42

Hunt, A., S. Moyer and T. Shevlin. 2000. Earnings volatility, earnings management and equity value. University of Washington working paper. Joos, P. and M. Lang. 1994. The effects of accounting diversity: evidence from the European Union. Journal of Accounting Research (Supplement) 32: 141-175. Kormendi, R. and R. Lipe. 1987. Earnings innovations, earnings persistence, and stock returns. Journal of Business (July): 323-345. Lee, C. 1999. Accounting-based valuation: impact on business practices and research. Accounting Horizons 13, 4: 413-425. Leuz, C., D. Nanda and P. Wisocki. 2003. Earnings management and investor protection: An international comparison. Journal of Financial Economics, forthcoming. Lev, B. 1983. Some economic determinants of the time-series properties of earnings. Journal of Accounting and Economics 5: 31-38. Lev, B. and P. Zarowin. 1999. The boundaries of financial reporting and how to extend them. Journal of Accounting Research 37, 2: 353-385. Lipe, R. 1990. The relation between stock returns and accounting earnings given alternative information. The Accounting Review 65,1 : 49-71. Penman, S. 2001. Financial Statement Analysis and Security Valuation. New York: McGraw Hill/Irwin. Penman, S. and X-J. Zhang. 2002. Accounting conservatism, the quality of earnings and stock returns. The Accounting Review 77, 2: 237-264. Pope, P. and M. Walker. 1999. International differences in the timeliness, conservatism and classification of earnings. Journal of Accounting Research (Supplement): 53-99. Ronen , J. and S. Sadan. 1981. Smoothing Income Numbers: Objectives, Means and Implications. Reading, MA: Addison Wesley. Thomas, J. and H. Zhang. 2002. Value-relevance properties of smoothed earnings. Columbia University and University of Illinois at Chicago working paper.