Accounting - Based Estimates of The Cost of Capital - A Third Way
Accounting - Based Estimates of The Cost of Capital - A Third Way
Accounting - Based Estimates of The Cost of Capital - A Third Way
Stephen Penman
Julie Zhu
Boston University and Shanghai Advanced Institute of Finance, Jiao Tong University
FIRST DRAFT
This paper offers an approach for estimating the cost of capital from observed accounting
information and compares the resulting estimates to so-called implied cost of capital (ICC)
calculations and those estimated with asset pricing models. The approach is based on two ideas.
First, buying expected earnings growth is risky; thus, any variable that predicts expected
earnings that is at risk of not being realized is potentially an indicator of the cost of capital.
Second, accounting principles induce earnings growth that ties to risk; thus, an accounting
number that results from this accounting is potentially an indicator the cost of capital. The cost of
capital estimates perform well in validation tests, in contrast to the alternatives that are the
current standards.
Accounting-Based Estimates of the Cost of Capital: A Third Way
1. Introduction
Considerable research in finance is devoted to developing models that deliver a measure of the
cost of capital (otherwise called the expected return, the discount rate, or the required return).
While the structure of no-arbitrage asset pricing models is understood as a matter of theory, the
endeavor has been frustrated by difficulties in identifying common factors, the risk premiums for
these factors, and the sensitivities of asset returns to the factors. In response, accounting
information has been brought to the task. One approach (in the vein of Claus and Thomas, 2001
and Gebhardt, Lee, and Swaminathan, 2001) estimates the so-called implied cost of capital (ICC)
as the internal rate of return that reconciles current price to earnings forecasts and a long-term
growth rate. A second approach (in the vein of Lyle, Callen, and Elliott, 2013) imputes the
estimate from models that make assumptions about the evolution of accounting numbers and
their connection to prices. This paper proposes a third accounting-based approach and compares
The approach applies the framework of Penman, Reggiani, Richardson, and Tuna (2015)
that connects expected stock returns to accounting numbers: The expected return (cost of capital)
is indicated by observables that forecast expected earnings and earnings growth at risk. Applying
the framework, Penman and Zhu (2014) show that selected accounting variables previously
associated with so-called “anomalous” returns exhibit the prescribed properties. Penman and
Yehuda (2015) show further that these accounting variables are contemporaneously priced in the
This paper picks up where these papers left off, leading to a cost-of-capital estimate. It
applies a formal process for identifying accounting variables that then combine into a cost-of-
1
capital estimate. First, the relevant variables connect a priori to earnings growth and risk under
accounting principles. Second, the variables not only predict earnings growth empirically but
also indicate the risk that growth will not be realized. Third, as validation, the identified variables
predict stock returns. The resultant cost-of-capital is then estimated from the identified variables
out of sample, with further validation that they forecast growth, risk, and returns (out-of-sample).
The paper then compares the estimate to alternatives, including ICC estimates and those
from asset pricing models. Our estimates exhibit characteristics that one expects of a valid cost-
of-capital estimate, but are quite different from the alternatives that are widely applied as the
returns and the risk in those returns rather than in-sample goodness-of-fit criteria. Forward
We find that the relevant accounting information is not incorporated in ICC estimates.
Growth expectations enter into the ICC, but we show both analytically and empirically how the
ICC calculation fails to capture the associated risk. While ICC estimates have had difficulty it
Cost-of-capital estimates from the Capital Asset Pricing Model (CAPM), implemented
with historical betas, have little relation to our estimates, nor do they predict actual returns.
However, our estimates predict forward betas, so the ability of CAPM estimates to predict actual
returns improves with updated beta estimates implied by the accounting information.
“characteristics” rather than “factors.” As it turns out, the Fama and French (1993) three-factor
model and Fama and French (2015a) five-factor extension incorporate some of the information
we identify as important, as does the investment model in Hou, Xue, and Zhang (2015a).
2
However, in the construction of “factor-mimicking portfolios” and the estimation of sensitivities
to those factors, this information is packaged in such a way to produce expected-return estimates
that correspond little to ours and which fail to validate against actual realized returns. We do find
that the sensitivity or returns to the Fama and French book-to-price factor is increasing in our
cost-of-capital estimate (as our analysis predicts it should), as does sensitivity (beta) to the
market factor. However, sensitivity to the other factors in the five-factor model and the Hou,
Xue, and Zhang (2015a) model vary little with our cost-of-capital estimate, even though those
factors are nominally based on similar accounting information. Our estimates indicate expected
returns that are not explained by these pricing models, and the amount of unexplained expected
The ICC approach warrants particular attention because it is applied extensively in accounting
research to answer the question: What is the effect of X on the cost-of-capital? (where X can be
more). The ICC has also been applied to validate asset pricing models, in Hou, Xue, and Zhang
(2015b), for example. The approach is applied in Claus and Thomas (2001), Gebhardt, Lee, and
Swaminathan (2001), Easton, Taylor, Shroff, and Sougiannis (2002), Easton (2004), and Gode
and Monhanram (2003), to name just a few.1 By and large, the approach has not been particularly
successful in validation tests, though more recent papers that add refinements show some
improvement.2 Most strikingly, the research has had difficulty in predicting average returns, a
1
For review and evaluation of this research, see Easton and Mohanran (2005, 2016), Botosan and Plumlee (2005),
Easton (2007), Botosan, Plumlee, and Wen (2011), and Echterling, Eierle, and Ketterer (2015).
2
See, for example, Huang, Natarajan, and Radhakrishnan (2006), Nekrasov and Ogneva (2011), Hou, van Dijk, and
Zhang (2012), Larocque (2013), Mohanram and Gode (2013), Ashton and Wang (2013), Fitzgerald, Gray, Hall, and
Jeyaraj (2013), Lee, So, and Wang (2015), and Li and Monhanram (2014).
3
property required of a valid estimate.3 This is quite remarkable given that many accounting
numbers (with less pretense to being the expected return) readily predict returns (and robustly
so), for example, earnings-to-price (E/P), book-to-price (B/P), accruals, and asset growth.
First, the cost of capital is inferred from accounting observables (in the present) rather
than analysts’ forecasts and/or assumptions about future growth rates. Analysts’ forecasts, said to
Second, in contrast to many ICC papers that specify the same growth rate for all firms,
expected growth varies over firms. Thus observables that indicate the cost of capital do so
Third, growth connects to risk. Rather than viewing the cost of capital, r, and the growth
rate, g, as separate inputs in estimating the ICC, our approach incorporates the idea, advanced in
Ohlson (2008) and Penman and Reggiani (2013) and supported by the evidence, that r and g are
related: Higher expected growth implies higher risk on average. In the typical reverse
engineering exercise that elicits the ICC, r ‒ g is an input in the so-called “terminal value” of the
valuation model assumed for the purpose. An increase in g with no effect on r has quite different
implications for price (from which the ICC is inferred) than a corresponding increase in g that
leaves r ‒ g (and price) unchanged because r increases along with g. While some ICC papers
(like Easton, Taylor, Shroff, and Sougiannis, 2002, and Nekrasov and Ogneva, 2011) allow for
variation of growth rates in estimating the ICC, our approach connects that variation to risk and
3
See Easton and Monahan (2005 and 2016), Guay, Kothari, and Shu (2011), and Botosan, Plumlee, and Wen
(2011), for example.
4
Hou, van Dijk, and Zhang (2012) substitute forecasts based on accounting observables for analysts’ forecasts but
maintain assumptions on growth rates for the long run.
4
the cost of capital. Thus, it renders quite different cost of capital estimates to these papers that
Fourth, the approach embeds accounting principles that govern the recognition of
earnings and tie expected earnings growth to risk. Accounting numbers that indicate the cost of
capital (in the first point above) are identified as numbers that indicate risky growth expectations,
both as a matter of accounting principle and with empirical support to that effect.
Fifth, while the ICC is estimated as a constant over all future time—a feature that is
Sixth, while the ICC has a circularity problem—estimates that employ price cannot be
used in valuations to challenge the price—our estimates, estimated and validated out of sample,
Seventh, the resulting cost of capital estimates predict returns robustly out of sample and
The key conceptual points that differentiate our approach from the ICC are the third point
that ties growth to risk and the recognition of the accounting principles that make the tie in the
fourth point. To sharpen the comparison of the approaches, the paper estimates the ICC and
growth rates under the Easton, Taylor, Shroff, and Sougiannis (2002) procedures and finds that,
not only do the cost of capital estimates differ significantly from ours, but the implied growth
rates that are jointly estimated bear little resemblance to the expected growth that indicates risk
and which is at the core of our approach. The procedures in Gebhardt, Lee, and Swaminathan
(2001) handle growth differently but the resultant cost-of-capital estimates also have little
5
correspondence with ours. That said, we do not carry out a comprehensive comparison against all
ICC measures, but many of these assume a constant growth rate over firms.
evolution of earnings. Assumed fixed parameters project future earnings from current earnings
and book values, and also imbed the discount rate (cost of capital). As prices are based on
expectations of future earnings with a discount for risk, the parameters also connect earnings and
book values to prices. Accordingly, the discount rate (the cost of capital) is implied by observed
accounting numbers, prices, and estimates of the parameters. In this vein, Lyle, Callen, and
Elliott (2013) assume the Ohlson (1995) “unbiased accounting” with autoregressive accounting
dynamics to yield an expression for the cost of capital, with parameters estimated in sample then
applied out of sample to produce cost of capital estimates. Christodoulou, Clubb, and McLeay
(2016) apply similar dynamics but without incorporating prices. Lyle and Wang (2015) embrace
the Voulteenaho (2002) tautology to describe the expected return in terms of log book-to-price
and expected log book return on equity (ROE) and then assume that log ROE evolves under an
AR(1) parameter such that ROE is expected to equal the cost of capital in the long run (and thus
market values equals book value in expectation). The cost of capital for future periods is then
Chattopadyhay, Lyle, and Wang (2015) apply a similar framework but allow expected ROE and
the cost of capital to differ by a constant in the long run (with book-to-market correspondingly
6
These approaches are based on observables rather than forecasts and allow for a changing
cost of capital. Thus, like the first and fifth points above, they contrast with the ICC approach.
And, like the seventh point, the estimated cost of capital does predict returns. However, the
papers are not in accord with the accounting they employ. The Voulteenaho (2002) tautology
introduces book-to-price and ROE in log form but, while Ohlson (1995) shows that GAAP book
values and earnings honor the Miller and Modigliani (1961) dividend irrelevance property so
foundational to valuation theory, this property is violated in log form; the evolution of log book-
to-market and log ROE is determined by payout. Further, the assumed parameterization is
critical. Though not often recognized, the parameters imbed accounting principles for measuring
earnings, but there is no explanation of the type of accounting implicitly assumed and how it
reveals risk. Nor is there an explanation that the accounting is representative of GAAP
growth, the feature in the second and third points above that distinguish our approach from ICC
estimation and also from these papers. An autoregressive assumption implies that, for the typical
case of price greater than book value, the premium of price relative to book value declines over
time. However, expected earnings growth (over that from retention) implies an increase in
accommodation of growth and its connection to risk in the second and third points that thus
distinguishes our paper from these papers as well as the ICC approach. In addition, the fourth
point connects GAAP accounting to growth and risk. Rather than governed by an assumed fixed
5
The autoregressive assumptions is usually explained by “the forces of competition” that are said to drive book rates
of return towards the cost of capital overtime, but there is no explicit recognition of the accounting that also
determines the book rate of return.
7
parameter, the accounting evolves over time with the resolution of risky growth expectations,
and it is this process that produces numbers that inform about the expected return.6
We contrast our approach with the parametric approaches here to be comprehensive, but
do not compare the estimates empirically in this paper. That is partly done in Penman and
Yehuda (2015) where the comparison is to the papers that assume the Voulteenaho (2002) model
with an autoregressive assumptions: Those papers produce expected return estimates that are at
however, that Lyle, Callen, and Elliott (2013) estimates of expected returns incorporate some of
the accounting variables underlying our estimate (though for different reasons) and are
1.3 A Caveat
Before proceeding we bring the standard caveat. In estimating the cost of capital, we employ
observed stock prices and validate against observed returns. Thus, we assume that market prices
are set efficiently to yield expected returns commensurate with risk, the assumption under which
As we pool data over a large set of firms and a long time period, it is only required that
prices are efficient on average. Further, our analysis explicitly connects expected returns to risk
and supplies validation with the data; observables that indicate the expected return connect to
risk a priori, and the resulting cost-of-capital estimates are associated with fundamental (non-
That said, without a benchmark of the “true” cost of capital, we in no way close the
market efficiency debate. One cannot deny alternative behavioral explanations for our findings
6
That said, our approach does not provide an explicit relation between accounting numbers and returns, an
admirable feature of these papers. On the other hand, that relation is by assumption and it is this assumption that is at
issue. Penman (2016) provides a detailed critique of the parametric approaches.
8
without further tests against those alternatives. The caveat remains: Despite evidence to the
contrary, we could be picking up expected returns due to mispricing. Nevertheless, the estimates
of expected returns that we produce are based on actual investor experience over many years.
They are persistently observed, indicating they are returns that investors should expect. If they
are persistent “abnormal returns,” then we have documented persistent mispricing, but we have
also documented that the so-called “abnormal returns” come with risk, bearing a further caveat.
We turn the approach for estimating the ICC around: Rather than estimating an ICC from current
prices and earnings forecasts and then asking (for validation) whether that estimate predicts stock
returns, we estimate the cost of capital by predicting stock returns directly. Validation is then
established out of sample. In a large sample representative of ex ante expectations, those returns
are, of average, the expected return to investing at the price that is the input to the ICC
calculation. They are also returns associated with the realization of the (analysts’) earnings
forecasts input to the ICC and thus, on average, the returns indicated by those expectations if
However, the accounting observables that predict those returns are identified, not by data
dredging to “see what works,” but under organizing principles that connect the accounting
The first organizing principle is from the framework of Penman, Reggiani, Richardson, and Tuna
(2015) that explicitly connects the expected stock return to accounting numbers: Given clean-
surplus accounting, the expected stock return for the forward year is given by the forward
earnings yield plus the price-denominated expected change in the premium of price over book
9
value during the forward year. That premium, in turn, is driven by expected earnings growth. The
benchmark case is that of no-growth where the expected change in premium is zero and the
expected return is equal to the expected earnings yield. Growth (over normal growth from
retention) implies an increase in the premium. Thus, any variable the predicts forward earnings
and subsequent earnings growth that investors deem to be at risk will capture the discount in
price for that risk and accordingly will indicate the expected return (the cost of capital). Penman
(2016) elaborates and the appendix to Penman, Reggiani, Richardson, and Tuna (2015)
The second organizing principle concerns the accounting that connects growth to risk. To
introduce the ideas, consider the book-to-price ratio (B/P) that features so prominently in asset
pricing models. A (presumably riskless) U.S. government bond fund has B/P = 1: One trades in
and out of the fund at book value (NAV). However a risky equity fund also has B/P = 1, so B/P
cannot differentiate risk and return in this case. The reason is the accounting: Mark-to-market
accounting (or fair value accounting, more generally) takes away the ability of B/P to indicate
risk. So, if B/P has anything to do with risk and the expected return in the more typical case of
B/P ≠ 1, it might have something to do with accounting that departs from fair value accounting.
The alternative accounting to fair value accounting is so-called historical cost accounting
which, apart from investment funds and some financial assets and liabilities, is pervasive. An
accounting principle, the so-called realization principle, drives historical cost accounting and
connects accounting numbers to risk and potentially the cost of capital: Book value increases
with the recognition of earnings but, under uncertainty, the recognition of earnings is deferred
until the uncertainty is resolved. Thus, the delay in the recognition of earnings that are expected
10
in the stock price indicates earnings at risk (of not being realized) while realized earnings reflect
The principle is applied in recognizing revenues only when the uncertainty about
receivable. In asset pricing terms, earnings are recognized only when the firm can book a low-
beta asset, cash or a near-cash (discounted) receivable. Until that point, revenue recognition is
deferred as higher beta, an expectation at risk of not being realized (the customer may not
materialize. Deferred earnings amount to expected earnings growth and the accounting thus
conservatism, assets are not booked (to book value) when earnings from the investments are
particularly uncertain; rather they are expensed against earnings immediately. R&D, advertising,
and promotion (brand building) are the classic examples, but the accounting is pervasive, applied
to investment to develop supply chains and distribution chains, employee training and retention,
software development, start-up and organization costs, and more. Conditional conservatism—
writing down booked assets on lowered expectations but refusing to recognize anticipated gains
until realized—reinforces this accounting upon arrival of updating information. The consequence
of this accounting is to reduce current earnings and induce higher subsequent earnings (growth)
that now attracts no expense amortization. But that growth, too, is tied to risk: The investments
are expensed because of uncertainty about outcomes.7 In contrast, investment that is booked to
the balance sheet (and does not reduce earnings) are deemed to be of lower risk; in contrast to
R&D, inventory and plant investments are made with a saleable product in view.
7
The accounting is explicitly tied to risk (or “the uncertainty of future benefits”) in FASB Statement No. 2 and IAS
38 on R&D accounting and IAS 37 on the recognition of contingent assets and liabilities.
11
This characterization is not conjecture; it is a description of how accounting works—
beginning “Accounting Principles” course. There is no imperative that the risk recognized by the
accounting principles is priced risk, of course. However, intuition suggests that stocks with
realization expected in the more distant future are more sensitive to shocks to the risk premiums.
Indeed, Chen (2016) takes this point to develop a consumption-based pricing model where
expected growth is priced with a higher expected return. The connection of the two accounting
principles to risk and the expected return has also been documented empirically. On the first
accounting principle, Penman and Reggiani (2013) find that the deferral of earnings recognition
(and the consequent higher expected earnings growth) forecasts how realized earnings differ
from expectation, and that risk is reflected in cross-sectional differences in stock returns. On the
second principle, Penman and Zhang (2015) document that conservative accounting also predicts
The application of these two organizing principles can be illustrated with a simple
valuation model with a constant cost of capital, r, and constant growth rate, g. Given full payout,
the valuation,
Et ( Earningst 1 )
Pt
rg
valuation (for a constant r) that is the foundation for all equity valuation.8 Thus,
8
The full payout assumption is unimportant. Payout (retention) other than full payout adds to earnings growth, g,
but does not add value under M&M conditions. The valuation isolates the growth that potentially affects price and
the expected return, r, and at the same time is M&M consistent. A constant cost of capital is, of course,
objectionable if no-arbitrage is implied, but the model here is just for illustration. Our approach lends itself to
estimating changing r.
12
Et ( Earningst 1 )
rg.
Pt
Et ( Earningst 1 )
In the case of no expected growth, r , the benchmark case in Penman,
Pt
Reggiani, Richardson, and Tuna (2015). The standard view is that expected growth, g, adds to
price and thus decreases the E/P ratio (and increases the P/E ratio). But, that is only so if r is held
constant while g varies. If r increases with g, one for one, r ‒ g does not change, nor does price;
expected growth is discounted in the price because it is risky. More generally, the first organizing
principle recognizes that E/P is not solely based on expected earnings growth, but also on the risk
The second organizing principle concerns the accounting for Earningst+1 in this model:
The deferral of earnings recognition and expensing risky investments reduces Earningst+1. For a
given Pt that embeds an expectation of total life-long earnings (for both t+1 and after), lower
Earningst+1 implies higher subsequent earnings and, on the lower base of Earningst+1, higher
earnings growth, g, in the denominator. If the growth so induced is pure accounting noise—it’s
growth, the growth adds to price. However, if the growth from earnings deferral is tied to risk,
the growth adds to r rather than price. If so, a higher g implies a higher r.
Et ( Earningst 1 )
Thus, a given E/P ratio, r g , can involve a high r with a high g or a
Pt
low r with a low g, and the investor is left with the task of finding information that discriminates.
The accounting that connects risk to growth has the potential to supply that information. A matrix
of nested portfolios in Exhibit 1 below illustrates how the relevant information might be elicited.
First, form portfolios on E/P ratios in the cross-section such that firms in each portfolio have the
13
same E/P and thus the same r ‒ g. Second, identify information which indicates earnings growth
under the accounting principle that connects that growth to risk. Third, within each E/P portfolio,
further sort firms into portfolios based on that information. Exhibit 1 displays a 5 × 5 set of
Exhibit 1. A Portfolio Formation Scheme that Identifies Expected Earnings Growth, g, that
is Related to Risk and the Cost of Capital, r.
LOW HIGH
E/P 2 3 4 E/P
2 ↓ ↓ ↓ ↓ ↓
Accounting
3 ↓ ↓ ↓ ↓ ↓
4 ↓ ↓ ↓ ↓ ↓
The top row of the exhibit gives the mean E/P ratios for five portfolios formed by ranking firms
on their E/P ratios. (These are actual mean trailing E/P ratios for U.S. traded stocks from
rankings every year, 1963-2012). As E/P is determined partly by r (and strictly so in the no-
growth case), these portfolios may indicate differential r, as the evidence indeed suggests.9
9
Considerable research, beginning with Basu (1977) shows that E/P predicts subsequent stock returns.
14
However, expected earnings growth also affects the E/P ratio. So, a vector of accounting
variables is identified that forecasts growth that connects to risk under accounting principles.
Firms are then sorted within E/P portfolios on this vector into five portfolios (down columns).
Et ( Earningst 1 )
With r g held constant in a given E/P portfolio, the information that sorts
Pt
on g is necessarily also a sort of firms on r (from low to high, as indicated in the matrix) if the
The identification of the vector requires that relevant accounting observables are
(i) those that a priori indicate growth after t+1 that is at risk under accounting principles,
(ii) are then confirmed empirically to predict that growth and the risk surrounding it, and
(iii) are also confirmed to predict returns, and in the same direction as they predict growth.
The requirement that variables predict growth and returns in the same direction is a consistency
condition: A variable that predicts higher (lower) returns must predict higher (lower) growth if it
is to convey risk under accounting principles. For this study, we limit the predictions to growth
To lead the selection of variables, we begin with book-to-price (B/P). This is a natural
starting point, for two reasons. First, earnings in the E/P ratio is the bottom-line number in the
income statement, and book value in B/P is the bottom-line (equity) number in the balance sheet.
Thus the implication of using these two aggregates is considered before a more detailed analysis
of their line-item components. Second (and more substantively), adding book value to earnings
captures the accounting principle that connects growth to risk. E/P × P/B = E/B ≡ ROE
15
(approximately).10 Thus, with reference to Exhibit 1, a ranking (down columns) on B/P for a
given E/P is an inverse ranking on ROE (return on equity), and ROE is affected by conservative
earnings and ROE (due to the immediate expensing) and increases expected earnings growth.
Further, with the investment now omitted from the balance sheet, the accounting yields high
ROE subsequently if earnings from the risky investment are realized (on a low book value base).
The if implies that the expected earnings growth is at risk, and the removal of the if on realization
implies the resolution of risk. In short, the accounting introduces a negative relation between
Accordingly, for a given E/P, a ranking on ROE down columns in Exhibit 1 (and a
corresponding ranking on B/P) identifies E/P ratios associated with low ROE where the earnings
in the E/P are reduced by deferrals and the expensing of investment, yielding higher g and
corresponding higher r if the risk identified by the accounting is priced risk. These E/P ratios are
distinguished from those associated with high ROE that result from the realization of earnings
from risky investments, implying lower risk and lower r and g. Penman and Zhang (2015)
elaborate and introduce measures of the effect of conservative accounting on ROE to show that,
for a given E/P, ROE is negatively related, not only to subsequent earnings growth in satisfaction
of requirement (ii), but also to the risk that the growth may not being realized. Further, ROE is
also negatively related to average returns in satisfaction of requirement (iii). Penman, Reggani,
Richardson, and Tuna (2015) report similar findings with B/P and thus explain, in part, the
10
The “approximate” qualification is because ROE is usually calculated on beginning (or average) book value for a
period whereas B/P involves end-of-period book value.
11
Feltham and Ohlson (1995) and Zhang (2000) model the effect of conservative accounting on ROE and growth,
albeit with no implication for risk or the cost of capital.
16
celebrated B/P effect in stock returns that is the basis for the Fama and French (1993) asset
information are selected under requirement (i), and are then confirmed to predict growth and
Earningsta2 E ( Earningst 1 ) B K
1 t 2 t j A j ut 2 (1)
Earningst 1 Pt Pt j 3
Et ( Earningst 1 ) Bt K
Rt 1 a b1 b2 b j Aj t 1 (2)
Pt Pt j 3
Earnings and prices are per share, and Earningsta 2 = Earningst 2 (r ft2 d t 1 ) where dt+1 is
dividend per share in t+1 and rft +2 is the yield on the one-year T bill for year t+2. The
reinvestment of dividends recognizes that dividends reduce earnings growth (or, alternatively
put, dividends can be reinvested to earn more earnings). The left-hand-side growth rate in
Earningsta 2 2
equation (1) is calculated as which ranges from -2.0 to +2.0. This
Earningsta 2 Earningst 1
measure produces a growth rate that is quite close to the standard measure with positive base
earnings in t+1 but accommodates the case where the base is negative, as well as compressing
The predictive regression (1) mirrors the portfolio scheme above. The regression holds
E/P constant in the first term, now at a point rather than for a portfolio. With E/P held constant,
variables selected under requirement (i) should jointly predict subsequent growth, and are
retained only if they do so. Regression (2) serves to test whether the selected variables connect to
the expected return: If those variables indicate priced risk, they should predict returns as well as
growth. The identified set of accounting observables that satisfy these requirements are thus
17
deemed to indicate the expected return one year ahead (for t+1). The approach can be adapted to
forecasts of growth and returns for t+2 and subsequent periods (subject to dealing with
survivorship issues).
Et ( Earningst 1 )
To cast the analysis in terms of accounting observables, is set equal to a
Pt
Earningst 12
forecast given by . For added accounting observables beyond Bt/Pt, Aj, j = 3,…, K,
Pt
we identify variables that indicate growth and risk under accounting principles and which predict
both future growth and returns in regressions (1) and (2) in Penman and Zhu (2014). These are
also variables that market contemporaneously prices as bearing expected-return news in Penman
and Yehuda (2015). This may not be the best or complete set, but our aim is to demonstrate an
approach rather elicit the definitive information set. We provide justification in the appendix as
to why identified variables are those that result from accounting principles that connect growth to
risk, in satisfaction of requirement (i). The appendix also reports that these variables satisfy
Significantly, some of these are characteristics that have been identified as return
predictors, either analytically or via data dredging, to construct return factors in asset pricing
models. Thus we are able to take our analysis to a comparison with models that incorporate
12
Earnings is before extraordinary items (Compustat item IB) and special items (item SPI), minus preferred
dividends (item DVP), with a tax allocation to special items at the prevailing Federal statutory corporate income tax
rate for the year. Earningst , so calculated, is strongly correlated in the cross-section with realized forward
Pt
earnings-to-price, Earningst 1 , with an average Spearman correlation of 0.63.
Pt
13
Summary statistics for these variables and test statistics from in-sample estimation of regressions (1) and (2) are in
Penman and Zhu (2014). Two of the variables in the appendix (EXTFIN and NSI) pertain to financing activities
which, while correlated with the other variables, are not affected by the accounting we have in mind. However, they
do introduce leverage effects on growth, risk, and the expected return. Penman and Yehuda (2015) obtain similar
expected return predictions when these variables are excluded, as do we in this paper.
18
similar information but with a different construction. Others have been identified as “anomalies”
(unexplained by these pricing models), but the appendix explains how they connect to growth
We produce cost of capital estimates for each year, 1981-2012, although the data period
runs from 1971 to 2014 covering estimation periods and periods for out-of-sample tests.14 For
each year, 1971-2012, regression equation (2) is estimated from the cross section with a
parsimonious set of accounting variables that satisfy requirement (i) and are validated to predict
growth in regression (1) in satisfaction of requirement (ii). Then, for each year, 1981-2012, an
expected return is estimated for each firm by applying the mean coefficients estimated over the
prior 10 years with regression (2) to the relevant accounting variables for that firm. The estimates
The analysis covers all U.S. firms which are available on Compustat files for any of the
years, 1971-2012, and which have stock price and returns for the corresponding years on CRSP
files. Financial firms (in SIC codes 6000-6999) are excluded because the accounting differs from
that depicted as relevant here (for non-financial firms), and so are utility firms (in SIC codes
4900-4949) where the accounting numbers are partially a result of regulation. Firms were deleted
for any year in which Compustat reports a missing number for book value of common equity,
income before extraordinary items, common shares outstanding, or total assets. Firms with
negative book value for common equity or a per-share value of less than 50 cents were also
eliminated. Prices (Pt in the denominator of the regressions above) were observed on CRSP three
months after each fiscal year, by which time the annual accounting numbers (for fiscal year t)
14
The period was limited to post 1970 due to lack of data for some variables. However, data were available back to
1963 for calculating E/P, B/P, Accruals and Growth in Net Operating Assets. Estimates for 1973-1980 with just
these variables produced similar findings to those reported in the paper.
19
should have been reported. Returns (Rt+1), also observed on CRSP, are annual buy-and-hold
annual returns after this date, calculated as compounded monthly returns with an accommodation
for non-survivors. Details of the calculation of accounting variables are in the appendix.
The empirical analysis begins with a documentation of how estimated expected returns connect
to risky growth expectations and how that risk manifests itself in investment returns. At this point
we refer to “expected returns” rather than the “cost of capital,” for reasons that will become
evident.
Each year, 1981-2012, firms are ranked on their out-of-sample estimated returns, ERt, and
formed into 10 portfolios. Panel A of Table 1 reports the mean (over years) of the mean ERt for
these portfolios.15 The column next to the mean ERt reports the mean (over years) of the forecast
of earnings growth two years ahead from applying regression (1) out-of-sample each year.
Portfolios 1 and 2 aside, the ER are increasing in the growth forecasts: The expected return
connects to expected earnings growth. Means of yearly median forecasts (not reported), though
slightly lower for each portfolio, exhibit the same pattern over portfolios.
Most of the remainder of the table reports the accounting characteristics from which the
ERt are inferred (again, the mean over years of yearly portfolio means). The first characteristic,
E/P, is the starting point for the analysis, as depicted in Exhibit 1. With the exception of the
extreme ER portfolios that are associated with loss firms, E/P is similar across portfolios 4 – 9.
As E/P = r – g for positive earnings firms, the issue, then, is whether ERt indicates whether these
15
As firms have different fiscal-year ends, the portfolio features in this table and Table 2 do not necessarily align in
calendar time, but represent all stocks. Results are similar with just December 31 fiscal-year firms where the features
refer to the same calendar period.
20
similar E/P ratios involve high r and g or low r and g. The ERt for the portfolios and related
growth expectations indicate an answer in the affirmative. Interestingly, negative E/P are
associated with both extreme high and low ERt portfolios so the ERt estimation potentially
distinguishes loss firms with high risk and expected return from those with low risk and expected
return.
The association between the other characteristics and ERt accord with the predictions in
the appendix. This, of course, is partly by construction as the variables are identified as those that
forecast growth in the estimation period; the numbers here just indicate that the relationships are
stable out of sample. B/P is positively related to ERt and the growth forecasts, in confirmation
that ROE, with the effects of conservative accounting, predicts growth and risk (as explained
above). Sales growth is considerably higher for the lower ERt portfolios, indicating that the
resolution of uncertainty with sales realizations projects lower expected returns, as accounting
associated with lower two-year-ahead earnings growth forecasts (except for portfolio 1). The
higher sales realizations in the low ERt portfolios are translated into higher operating profit
margins in the table; sales growth realizations flow through to higher realized earnings. In
contrast, high ERt portfolios are those where sales growth and margin growth are relatively not
(yet) realized. Accruals, realized investments, and NOA growth exhibit the predicted patterns in
the appendix with respect to both ERt and forecasted growth. For example, the relatively high
investment for the low ER portfolios indicate that uncertain investment opportunities have been
realized—investment options have been exercised—and booked to the balance sheet rather than
expensed under conservative accounting as particular risky; the booked investment indicates that
expected earnings growth is more likely. Higher investment, along with higher sales growth and
21
profit margins, is associated in the table with higher financing (EXTFIN and NSI) that also
exhibit the predicted association with subsequent growth and expected returns in the appendix.16
The characteristics associated with loss firms (in portfolios 1, 2, 10, and, to a lesser
degree, portfolio 9) highlight our organizing principles. Loss firms in portfolios 1 and 2, with
substantial realized sales growth, increasing profit margins, and growing investment in
anticipation of more realizable sales and earnings in the future, are (intuitively) less risky than
those making losses in portfolio 10 with flat or declining sales and profit margins and no
investment growth. The loss firms in the low ERt portfolios are also growing their balance sheets
The final column in Panel A reports the mean expected return, ERt+1, re-estimated at the
end of year t+1 for the same firms that are in the relevant portfolio for ERt. The expected returns
are quite persistent, though exhibit some reversion to the mean in the extremes, particularly at
the lower end. The mean reversion is expected as firms become similar to the average over time
but may also be due to measurement error: Extreme ERt may be due in part to over- or under-
estimated expected returns—points that will be pertinent later in the paper. In unreported results,
the accounting characteristics for ERt portfolios in the table are also persistent; their t+1 values
have a similar pattern over ERt portfolios as their time-t values. For example, the negative
relationship between sales growth and ERt is also observed in their time-t+1 values, and so for
the other characteristics. Thus the returns are not associated with reversals (of accruals, for
example) that might be expected if the actual return difference over portfolios were due to
market mispricing. That, along with the persistence of the relationships over the sample period,
16
Leverage (net debt/market value of equity), not reported in the table, is decreasing in ER t. A negative correlation
is consistently observed between leverage and average returns when operating risk is not controlled for. See
Penman, Richardson, and Tuna (2007).
22
points to the accounting characteristics indicating return for risk rather than abnormal returns
Panel B of Table 1 reports the accounting payoffs (after time-t when ERt is estimated) for
the respective ERt portfolios. It begins with the forecast of two-year ahead earnings growth as in
Panel A, but now with that forecast compared to the actual (realized) earnings growth. With the
exception of portfolio 1, actual growth rates align over ERt portfolios, as they do with the
forecasted growth rates.17 Further, higher growth forecasts are associated with a higher standard
deviation (STD) and interdecile range for the actual growth rates, as also reported in Penman and
Zhu 2014: Portfolios with higher expected growth also have higher risk of actual growth
deviating from expectation to a larger degree, with higher probability of realizations in the tails
of the distribution—and these are firms with higher estimated ERt18. The STD and IDR measures
refer to the variation in means for portfolios over time so, to the extent risk is diversified in these
portfolios, the differences across portfolios reflect common risk. The four-year aggregate actual
earnings yields (four-years of earnings, t+1 to t+4, relative to price at t) are increasing in ERt; A
higher ERt, denominated in price, is rewarded with higher earnings denominated in the same
price, on average. But, again, that higher yield is at risk, as measured by the standard deviation
17
The comparison of expected and actual earnings growth for portfolio 1 suggests that the expected growth contains
considerable measurement error, and so for portfolio 10.
18
Realized growth two years ahead are affected by realized investment one year ahead. However, results are similar
for growth in residual earnings, with book value at the beginning of the year charged at the prevailing risk-free rate
for the year. This also deals with the issue that, for a given E/P in the Exhibit 1 construction, growth may differ over
portfolios because of different payout (retention).
23
Similar patterns for the standard deviation and IDR are observed for realized one-year
ahead earnings, EPSt+1/Pt.19 These results are not reported, but the table does report portfolio
earnings betas for t+1. The betas are estimated from regressions of the time-series of actual
(market-wide) earnings relative to price. Betas for the sensitivity of earnings changes in t+1 to
changes in market-wide earnings are also given. These are betas actually experienced during the
forward year that ERt forecasts. Higher (lower) ERt forecasts higher (lower) betas; ERt predicts
In summary, while Panel A of Table 1 indicates that ERt is related to earnings growth
expectations, Panels B indicates that ERt is also related to the risk around those expectations.
There is some reservations about portfolio 1, though one must be skeptical about means taken
over the extremes: Observations with high measurement error are in the extremes.
A question remains open: Does ERt imbed a priced premium for the risk we have
observed?
In answer to this question, Table 2 reports the realized return outcomes for the ten ERt portfolios
and the risk incurred with those realizations. The various metrics are calculated from the time
series of actual (realized) annual portfolio returns for each t+1 year over the sample period, the
returns that ERt forecasts. Mean and median actual returns are almost monotonically increasing
in ERt and indeed are quite similar to the ERt estimates. The correspondence is quite impressive
19
The results in the table are subject to any survivor bias. The percentage of firms not surviving for two years (due
to liquidation or takeover) is (in percent) 6.69, 5.27, 2.93, 2.31, 1.97, 1.74, 2.01, 1.91, 1.97, and 2.6 for portfolios 1 -
10 respectively. The similar pattern in the standard deviation and IDR for actual EPS t+1/Pt (where the non-survivor
issue is reduced) is thus reassuring.
24
for out-of-sample estimates. It is similar for the last half of the sample period as the first half,
indicating persistence in the relationship. The standard deviation of actual returns (again,
calculated for portfolios over time) increases in the high ER portfolios, but otherwise the
differences are small, and portfolio 1 (again) has a similar standard deviation to portfolio 9. A
comparison of mean returns to the standard deviation reveals that Sharpe ratios are increasing
across portfolios. However, this higher return per unit of standard deviation for higher ERt is
associated with an increase in the range of return outcomes. And the kurtosis measures indicate
that there is increasingly more realizations in the tails as ERt increases, while the relative
skewness measures indicate that higher ERt portfolios yield compensation for the risk on the
upside. It appears that the accounting information indicates the likelihood of returns in the
extremes, and it is the extremes with which investors are particularly concerned.
The predictable average realized returns in the table indicate return for risk if the risk is
priced systematic risk. So, the table also reports beta sensitivities to common return factors
appearing in asset pricing models. The forward market (CAPM) betas—estimated from the
annual actual t+1 portfolio returns regressed on the return for the market portfolio over time—
complement the earnings betas in Table1, Panel B. They are higher for high ERt portfolios,
though portfolio1 reports a high beta. Betas in up-markets (years when the value-weighted CRSP
index return was greater than 10%) are also increasing in ERt, as are down-market betas (years
when the value-weighted CRSP index return was less than -10%). While the portfolio 1 up-
market beta is 1.77, the down-market beta is only 0.48; it appears that this portfolio provides a
hedge against bad times when wealth of investors is low, a property that implies lower risk and a
25
lower expected return in the Merton (1973) intertemporal asset pricing model.20 Note that these
betas are those actually experienced during the t+1 year, not historical betas.21
Table 2 also reports the sensitivity of portfolio returns to the additional return factors
(over the market factor) that appear in the Fama and French (2015) five-factor model, HML
(book-to-price), SMB (size), RMW (profitability), and CMA (investment) factors. We also
estimated the sensitivity of the returns to the factors in the Hou, Xue, and Zhang (2015a) (HXZ)
so-called Investment CAPM. The results for the latter’s size (ME) and ROE factors were similar
to those for the corresponding to the FF SMB and RMW factors, so we only report the results for
the investment factor, I/A. We present these findings more tentatively, for reasons that will
become apparent when we evaluate ERt against expected returns from these models later. The
reported betas are estimated from time series regressions of actual excess monthly returns for the
portfolios in year t+1 on these factor returns and the excess market return. So, like the CAPM
As reported in Panel A of Table 1, ERt is positively related to B/P and here the sensitivity
of portfolios returns to the HML factor increases with ERt. While each portfolio is quite sensitive
to the SMB size factor—t-stats are all over 12.0—there is little difference in the SMB betas.
Unreported results indicate that size varies little over portfolios; the sensitivity to risk associated
20
Ang, Chen, and Xing (2006) model a premium for downside beta relative to upside beta and show empirically that
relative downside risk earns a return premium.
21
We did find some evidence that contradicts the beta results. We calculated mean daily returns for portfolios during
the forward year, separating days with pre-scheduled macro news announcements from non-announcement days.
There were 1,302 news days in the sample period, covering announcements of CPI and PPI revisions, employment
statistics, and interest rate decisions from the FOMC. Mean returns are increasing in ER t for the non-announcement
days. However, while the mean returns during announcement days were significantly higher than non-announcement
days for all portfolios, they were similar across portfolios; the higher ERt firms did not earn higher average returns
to compensate for risk of macro news.
26
with size (incremental to other factors) is within portfolios rather than across portfolios.22 The
betas on the profitability factor, RMW, are all negative and those on the CMA investment factor
are not significantly different from zero (with the exception of portfolio 1). The betas on the
investment factor, I/A, are positively related to ERt. With investment negatively related to ERt in
Panel of Table 1 and also negatively related to investment in the Investment CAPM, the
alignment is agreeable.
With respect to RMW and CMA (and the ROE in the HXZ model), the findings could be
interpreted as ERt not identifying the risk in these factors, but only if one accepts that these
factors actually are risk factors. As the factors in FF are largely generated by data search one
must remain skeptical, particularly as ERt is based in part on ROE and investment. After all, ERt
is related to actual mean returns in the table, whereas these factor sensitivities are not. The
question can be turned around: Why does sensitivity to these factors not indicate the expected
returns, actual returns, and risk indicated by ERt? We return to this issue later.
The cross-sectional variation in portfolio ERt is evident from these tables. Within ERt
portfolios, the standard deviation of the yearly estimates over time is close to 0.04 for most
portfolios, except for portfolio 1 where it is 0.056. We do not have a benchmark for how ERt
should vary over time (with changes in interest rates and risk premia), but this variation is not
large enough to suggest that the estimates are wildly fluctuating. Nevertheless, some of the
variation is presumably due to estimation error. Figure 1 plots the variation of the 10-year
moving average of the estimates over time, for 1990-2012. The averaging washes out mean-zero
22
Size as measured by ln(market capitalization) is slightly lower for portfolio 10, 4.2 versus the typical size of about
5.4 for other portfolios. Repeating the analysis with value-weighted returns for ERt portfolios, the SMB sensitivity
coefficients were considerably lower, with an average of 0.18 across portfolios (but still varying little across
portfolios). Thus the betas on SMB in Table 2 are due to the presence of small firms in all portfolios. Similar
findings on HML and SMB were observed with the original three-factor Fama and French model. Adding the UMD
momentum factor to these three factors (as in some extensions of the model), betas on this factor are consistently
negative over portfolios.
27
estimation error if the “true” expected return is changing slowly. It is clear that the differences in
ERt between portfolios is maintained over time. Further, the estimates co-vary over time,
suggesting that they contain a common varying systematic risk premium. For all portfolios, the
estimates are lower in the mid-to-late 1990s, a period of quite high ex post returns.23 But, relative
to the variation across portfolios, the year-to-year estimates are quite stable, as are also
Research has uncovered a large number of characteristics that predict returns, but far too many to
suspect that they independently do so.24 Most of these have been “discovered” by data dredging.
There is a need to reduce the dimensionality, not only to exclude those that just load with a
“significant t-statistic” by luck in the data dredging, but also to produce a parsimonious set of
legitimate predictors from the many correlated variables. Lewellen (2015) does so, with out-of-
sample validation, but as a statistical exercise that isolates predictors deemed “insignificant”
given other included variables. Most of the relevant variables are accounting variables. In
contrast to the statistical approach, this paper reduces the dimensionality by filtering variables
into the analysis a priori on the basis of organizing principles that connect these accounting
variables to risk.
In the Lewellen (2015) paper, book-to-market, size, and momentum explain a good deal
of the cross-sectional variation in the out-of-sample expected returns estimated from the full set
of predictors, as they do actual returns. Variables identified by our analysis, such as accruals,
23
The seemingly low ERt estimates for portfolio 1 in Table 1 are due mainly to this period (when they were
negative). Some claim, without much justification, that the mid-to-late 1990s was a period when risk premiums
declined.
24
Harvey, Liu, and Zhu (2016) find 316 predictors, a number they say likely under-represents the total. Green,
Hand, and Zhang (2013 and 2014) find that that, of 333 characteristics that have been reported as predictors of stock
returns, many predict returns incrementally to each other.
28
asset growth, and profitability, add significantly less explanation, while a number of other
attributes not entertained by us (and largely not accounting variables) add very little.
Table 3 draws a contrast. It reports results of regressions of forward actual returns, Rt+1,
on out-of-sample ERt, with and without variables identified as relatively important in the
Lewellen paper. The regressions are estimated in the cross section each year with mean
coefficients over years and t-statistics on the means (Fama-Macbeth style) reported in the table.
The regression at the portfolio level with ERt alone produces a mean R2 of 0.55, as might be
expected from the alignment of ERt with actual t+1 returns in Table 2. The R2 is much reduced
when the regressions are run at the individual firm level, of course, but the t-statistic on ERt is
6.12.25 Remarkably, the mean coefficient estimate of 0.979 is very close to 1.0, with a standard
error of 0.159 and a t-statistic for the mean relative to 1.0 of -0.17. The corresponding t-statistic
on the mean slope coefficient of 1.065 for the portfolio regressions is 0.37. In both cases, the
mean intercept (average bias) is not significantly different from zero. The slopes compare with a
slope of 0.76 in the Lewellen paper that is significantly different from 1.0, with a standard error
of 0.13.26 The slope of 1.0 takes on further significance with the understanding that it is an
estimate of the return on a minimum-variance long-short portfolio with unit net exposure to ERt ,
compared with a similar figure in Lewellen (2015). The slopes are higher in the early 1990s and
early 2000s (periods of lower actual returns) and lower in the late 1990s (a period of high actual
returns).
25
It is difficult to benchmark the R2. Most papers that predict returns report R2 from in-sample estimation (typically
about 0.015) whereas this is out of sample. The mean R2 from estimating regression (2) in sample is 0.045.
26
Lewellen (2015) also reports a mean slope of 0.74 with monthly return regressions, with a standard error of 0.07.
29
In Table 3, the coefficient on ERt changes little with the addition of the variables in the
Lewellen paper that are important to explaining returns, and none of these loads with a
significant t-statistic.27 Of course, they may be collinear so may add explanatory power jointly,
and the R2 is 0.050 compared with the 0.015 with ERt alone. However, the added variables are
fitting in-sample (where the R2 informs about the amount of contemporaneous volatility
Many estimates of the ICC assume a constant growth rate across firms. We limit our comparison
to two that allow for differing growth rates. The first is the ICC in Gephardt, Lee, and
Swaminathan (2001) (GLS) where a growth rate is implied by assuming the reversion of ROE to
an industry average. The second is that in Easton, Taylor, Shroff, and Sougiannis (2002) (ETSS)
Panel A of Table 4 reports the respective mean ICC for firms in ERt portfolios. The GLS
estimates vary little over the ERt portfolios, though they increase slightly as ERt increases. As in
the original paper, the ETSS estimates are inferred from cross-sectional regressions of expected
four-year ROE on Pt/Bt for the portfolios, with the four-years of expected earnings for the ROE
27
We also added all the variables in Panel A of Table 1 to ERt to assess whether the aggregate ERt captured all the
information in the variables underlying its construction. Only ΔNOAt loaded with a significant mean coefficient
(with a t-statistic of -2.14).
28
One could argue that the explanatory power of ERt is not quite out-of-sample given that the predicative ability of
some of the variables that enter the calculation of ERt have been observed in previous studies in some of the out-of-
sample periods here. However, ERt is not simply identified on the basis of in-sample correlations; rather, it is based
on a filter that requires an a priori connection to risk and the requirement that variables also fit in the growth
regression (1) before entering regression (2). Further, the t-statistics on ERt in Table 3 are much higher than the 3.0
that adjusts for multiple comparisons in Harvey, Liu, and Zhu (2016). The correlations previously discovered were
done so before 2000 (and many earlier) and Table 3 results hold after that date: The mean slope for the individual
firm regression from 2000-2012 is 0.962.
30
calculation given by analysts’ forecasts.29 The mean estimated intercepts and slope coefficients
from these annual regressions are reported in the table. The ICC inferred from these estimates (r
in the table) are actually decreasing in ERt, though they vary little. Further, the implied ETSS
annualized growth rates (g in the table) are also decreasing over the ERt portfolios, in contrast to
Table 1 where both expected and actual earnings growth is increasing in ERt. The positive
relationship between r and g that our analysis posits and confirms is not at all evident in the
ETSS estimates. While there is a positive relationship between the ICC and the ETSS implied
growth rates, both are negatively related to ERt and to the forecasted and actual earnings growth
What could explain this? The reliance on analysts’ forecasts is a recognized problem, but
we think the issues go deeper. First, the ETSS growth rate is the growth rate in residual earnings
whereas we refer to growth in earnings. Residual earnings involves the cost of capital, so the
ETSS growth rate is itself a function of the cost of capital, a confusion of constructs. Second,
ETSS infer a constant growth rate for the long term on a base of four years of earnings forecasts
estimated ETSS with just one year of analysts’ forecasts (for positive earnings forecasts only as
ETSS does not work for negative earnings), and obtained similar results to those in Table 4—
both portfolio r and g are negatively related to ERt.30 Nevertheless, the estimated g is still a
constant growth rate for all periods in the future, constraining the two-year ahead growth rate to
29
Our analysis is for 1981-2012. To ensure consistency with the ETSS findings, we first replicated their analysis for
their sample period up to 1998. We maintained their criteria for dealing with data issues and reinvestment rates for
the longer period. We also obtained similar results when we used analysts’ forecasts and P/B ratios three months
after fiscal-year end rather than at fiscal-year end, when we used IBES prices and shares outstanding rather than
those from Compustat, when we made adjustments for differences in IBES and Compustat numbers for shares
outstanding, and when we used different dividend reinvestment rates.
30
Portfolio r decline from 10.3% for ERt portfolio 1 to 7.9% for portfolio 10, and g declines from 9.2% for portfolio
1 to 3.3% for portfolio 10. Results are similar for growth after two years of analysts’ forecasts.
31
be the same as that is all subsequent years. One might expect that, for a given Pt, a higher growth
rate two years ahead would imply a lower growth rate in subsequent years (as the growth rate
into the conflicting estimates, we restate the ETSS formulation, starting from the version of the
residual earnings model where the constant growth rate starts two-years ahead (just for
simplicity):
Et ( Earningst 1 ) r.Bt
Pt Bt .
rg
Inverting,
B B
r t Et ( ROE t 1 ) (1 t ) g (3)
Pt Pt
(provided that ROEt+1 > g). That is, r is a weighted average of forward ROE and subsequent
growth where the weight is given by observed Bt/Pt. Dividing through by Pt/Bt and rearranging,
Pt
Et ( ROE t 1 ) g ( r g ) (4)
Bt
Pt
0 1
Bt
This is the equation that ETSS estimate in portfolios to extract γ0 = g and γ1 = r – g, with an
added error term because r and g are not likely to be the same for all stocks in a portfolio. An
alternative expression can also be derived that delivers the same r and g. Restating equation (3),
Et ( Earningst 1 ) B
r (1 t ) g
Pt Pt
and rearranging,
32
Et ( Earningst 1 ) B
(r g ) g t . (5)
Pt Pt
Thus, estimating r and g from the ETTS expression is equivalent to estimating them from a
Expression (5) connects E/P and B/P in a very different way to our setup. It depicts E/P
and B/P as linearly related, varying by a constant, g. In contrast, our formulation, depicted in
Exhibit 1, sees r and g both varying with B/P while holding E/P constant. That is so empirically:
Penman, Reggiani, Richardson and Tuna (2015) show explicitly how B/P orders growth and
average returns for a given E/P. Panel A of Table 1 shows that B/P varies over a wide range of
ERt portfolios where E/P is relatively constant, and that variation aligns with variation in the
growth rate as well as ERt. Also, negative E/P are associated with both high and low B/P,
inconsistent with the linear relationship in equation (5), and B/P also sorts on growth and the
expected return in this case. This contrasts with sorting on E/P and B/P pairs to infer r and g as in
equation (5).
A similar critique applies to equation (4) which ETSS apply to estimate r and g. By
Et ( Earningst 1 )
dividing the simple model that is our starting point (in Exhibit 1), rg,
Pt
through by Et(ROEt+1),
Bt rg
(6)
Pt Et ( ROE t 1 )
and
Pt
Et ( ROE t 1 ) ( r g ) (7)
Bt
33
This expression looks very much like the ETTS equation (4) except that g is now expected
growth in earnings rather than growth in residual earnings. Equation (6) simply says that, for a
Et ( Earningst 1 )
given r g , a lower Et(ROEt+1) yields a higher B/P, as recognized earlier in
Pt
connecting B/P to ROE. Introducing conservative accounting that reduces Et(ROEt+1) but also
increases g, it must be that, holding r – g and thus E/P constant, r also increases with the increase
Pt
in g and B/P. That is, both r and g change as Et(ROEt+1) and in equation (7) change. We see in
Bt
Pt
Table 1 that increases over ERt portfolios, along with expected and actual earnings growth,
Bt
and ERt validates against actual realized returns. Equation (7) and equation (4) are a basis for
Pt
estimation if higher g implies higher , which is the case when growth adds to price but not to
Bt
risk, but are not appropriate when growth adds to risk and the expected return such that r – g is
Panel B of Table 4 forms portfolios with cost-of-capital estimates from the Capital Asset Pricing
Model (CAPM) and a three-factor Fama and French model (FF). All estimates are calculated
with sensitivity (beta) coefficients estimated over prior 60-month periods, though results are not
sensitive to the length of the estimation period. For the CAPM, the historical market beta is
applied to a market risk premium of 5 percent for all stocks which, added to the prevailing risk-
free rate, yields the cost of capital. For the FF estimates, beta coefficients estimated on monthly
MKT, SMB, and HML factors over the prior 60 months are applied to mean factors returns over
the full sample period which, together with the estimated intercept, yield a cost-of-capital
estimate (annualized in the table). With the same factor risk premiums for each firm, differences
34
in CAPM and FF cost of capital estimates across firms are determined by the firm-specific
The CAPM betas are a good indication of the experiential (forward) betas (Betat+1 in the
table) experienced over the subsequent year when actual returns for the respective portfolios
were realized. However, those actual returns for t+1are similar over all portfolios, despite this
beta relationship. This contrasts with the relationship between ERt and actual returns in Table 2.
The table further reports that the CAPM cost-of-capital estimates are similar over ERt portfolios,
The Fama and French (FF) cost-of-capital estimates for portfolios are not related to actual
t+1 returns either, nor do the estimates vary much over ERt portfolios at the far right of the table.
The last two columns report intercepts and their t-statistics from time series regressions of ERt
portfolio excess monthly returns in actual return periods on the three factor returns. These
intercepts are returns unexplained by the FF model. They are increasing in ERt, indicating that
ERt captures aspects of the expected return not recognized in these pricing models.
Results are similar with expected return estimates from the FF five-factor model (that
includes the RMW and CMA factors investigated in Table 2), and those from the Hou, Xue, and
Zhang (2015a) Investment CAPM (that includes the investment factor in Table 2); cost of capital
estimates with these models vary little over ERt portfolios and are not related to actual forward
returns. Indeed, setting the expected return equal to the mean return over the past 60 months
31
Rather than using historical factor sensitivities to estimate the FF cost of capital, we estimated them for ER t
portfolios by applying the factor betas (for a three-factor model) estimated from actual returns during all t+1 years in
the sample period for those portfolios. We then applied those betas to the mean factor returns over the sample
period. The resulting estimates for ERt portfolios 1 – 10 (which are in sample) were 0.046, 0.104, 0.135, 0.157,
0.172, 0.167, 0.189, 0.176, 0.178, and 0.177, exhibiting some, but not a lot, of variation over portfolios. Shrinkage
35
The failure of these models to predict actual returns is well known.32 The goal here is to
see the extent to which they reflect the information in ERt and vice versa. With respect to the
CAPM estimates, the table informs that the historical beta is a poor indicator of actual mean
returns, whereas Table 2 indicates that ERt forecasts actual returns and forward return betas and
fundamental betas that align with those returns. Thus, it may not be that beta is a poor measure of
the risk, but that it varies overtime. An historical beta is not a good estimate of the forward
(experiential) beta, but the accounting information provides an update of the beta. Indeed,
Penman and Yehuda (2015) show that changes in ERt, calculated in a similar way as in this
paper, coincide with changes in beta. Thus, the results here are not to be interpreted as saying
that beta or the (conditional) CAPM is unimportant, but that accounting information can be
elicited to indicate the expected forward beta that investors will experience (and update the
historical beta).33
Stationarity of factor betas (and possibly the risk premiums on the factors) is also an issue
with FF models, as Fama and French (1997) recognize, and it may well be that accounting
earnings are deferred or realized, indicating changes in expected returns that cannot be captured
of beta coefficients estimated from past data may improve the results, but the analysis in Fama and French (1997)
does not hold out much hope.
32
See, for example, Fama and French (1997), Simin (2008), and Chattopadyhay, Lyle, and Wang (2015). The
failure is often attributed to non-stationarity in betas or factor risk premiums.
33
Estimating the CAPM cost-of-capital with the forward beta estimates reported in Table 2 for ER t portfolios yields
higher estimates for ERt portfolios 1, and 7 – 10, but little difference over portfolios 2 - 6. Applying the accounting
information underlying ERt to forecast beta changes is a potential alternative to the statistical approach to shrinking
betas in Blume (1975) and Vasicek (1973).
36
However there is another issue that differentiates our estimates and the FF estimates:
Some information is common to both approaches but the packaging of that information into an
expected-return estimate differs. The organizing principle for the construction of the FF-type
models involves searching for correlations in the data from cross-sectional (Fama and Macbeth)
regressions of returns of characteristics, then constructing return factors from the discovered
characteristics. B/P enters in this way as a basis for the HML factor. In contrast, B/P enters under
our organizing principles as an indicator of the expected returns because, under accounting
principles for a given E/P, it captures ROE and its connection, via conservative accounting, to
growth and risk. Penman and Zhang (2015) elaborates. ROE underlies the RMW factor, with
higher ROE indicating higher risk (it is said). But, under our organizing principles, ROE
indicates growth, risk, and the expected return negatively conditional on E/P. E/P = ROE × B/P
so, with B/P already a characteristic in the FF framework, ROE is capturing E/P, the starting
point in our approach rather than an incremental factor (and a factor missing from the original
three-factor FF model). Penman and Zhang (2015) show how the comparative statics that Fama
and French (2015a) employ to introduce the ROE factor are at odds with how accounting works.
And so with their identification of the investment factor. While investment enters in both our
approach and in the construction of the CMA factor in FF, the returns for ERt portfolios are not
information a priori and the packaging of that information into a cost-of-capital estimate. The
two approaches differ significantly in both aspects.34 The packing aspect is pertinent in light of
34
That is not to imply that we have identified all the relevant information or the appropriate way to combine the
components. Our cross-sectional regression approach imposes a linear relation between returns and characteristics,
and the estimation equally weights firms so that smaller firms have more representation than in the investment
universe. Again, the purpose of the paper is to demonstrate an approach rather than perfect it. The construction of
37
the observations that B/P, investment etc. predict returns but a FF cost of capital based on these
inputs does not. From this view, the sensitivities of ERt returns to FF factors in Table 2 are more
a commentary on the FF model than a demonstration of risk and return. The criterion we are
invoking is out-of-sample fit against the actual returns to investors in the cross-section, in
contrast to much of asset pricing research where the criterion is explaining variation in returns
over time with constructed factors and the significance of in-sample intercepts. That said, we do
not observe the “true” cost of capital for benchmarking, so all is relative.
constructing factors on the basis of selected characteristic return correlations, the resultant
models are evaluated on how well they explain or “digest” other “anomalies,” as in in Fama and
French (2015b), Hou, Xue, and Zhang (2015a), and Zhang (2015). In our construction, so-called
anomalies involving accounting variables are directly digested into the cost-of-capital estimate
under a priori criteria that connects them to risk. (That said, we have not entertained all
nominated anomalies.)35
The packing of information into an expected return estimate is also an issue with ICC
estimates. For example, the ETSS estimates recognize the same ROE and P/B as the FF five-
factor rendition as pertinent information, then tie them together in a linear relationship to extract
a cost-of-capital estimate. As discussed above, with E/P = E/B × B/P, that is a very different
handling of this information than in our construction. Lyle and Wang (2015) also embrace these
factors from portfolios in FF-type models relaxes the strict linearity, although there appears to be a continuing
unresolved discussion about the appropriate split points for forming portfolios, with results sensitive to these split
points. Daniel and Titman (1997) argue that characteristics rather than risk factor exposures are a better measure of
cross-sectional dispersion of returns.
35
Stambaugh and Yuan (2015) add two factors to the market and size factors in the FF models based on 11 anomaly
variables that include four of our variables in the appendix (and two others fairly close to them), and find that the
resultant models accommodate a large set on anomalies relative to the Fama and French (2015b) and Hou, Xue, and
Zhang (2015a). They choose to refer to the added factors as “mispricing factors.”
38
two pieces of information (in log form), but connect them to expected returns (with assumed
The analysis to this point demonstrates that information elicited from financial statements
informs about risk and the expected return for investing. However, there is a problem: The
estimated expected returns for the portfolios in Table 1, 17.4 percent at the median and 26.4
percent for portfolio 10, seem too high relative to what one expects of a cost-of-capital measure.
A comparison to the median CAPM cost-of-capital estimates in Table 4 makes the point (as does
The issue is one of sample bias: the mean and median actual returns in Table 2 also seem
out of line with perceived expected returns and the mean intercepts in Table 3 are close to zero.
This was a period when stocks paid off very well, on average, resulting in high actual returns and
high expected returns fitted to those actual returns. The range and kurtosis measures in Table 2
suggest that the accounting measures indicate susceptibility to both the extreme upside and
extreme downside. But the upside has dominated in this period, yielding the positive skewness in
payoffs for the high ERt portfolios. From our perspective, this makes sense: The risk of investing
is in buying growth, with down-side risk compensated with upside potential, and this was a
period when a bet on growth paid of well in the U.S.36 It was, after all, “the American Century.”
But that introduces the “Peso problem” that also confounds historical estimates of the equity risk
premium.
Converting the ERt estimates to an ex ante cost of capital requires adjustment for this ex
post bias. As the mean slope coefficients for actual returns regressed on ERt in Table 3 are almost
36
The sample bias is in all papers that estimate expected returns using realized return over the last 50 years or so.
See, for example, Lyle, Callen, and Elliott (2013). Lyle and Wang (2015), and Lewellen (2015).
39
exactly 1.0, the issue likely lies with the estimated intercept in estimating ERt with regression
First, regression (2) was re-estimated each year with the intercept constrained to be the
annual yield on the 10-year bond, Rft, for the relevant year. The explanatory variables now
explain the variation of returns in excess of the risk-free rate. Then, in fitting the ERt out-of-
sample, the intercept was set as the corresponding Rf for that year (with the mean rolling
While this procedure explains variation around the risk-free rate, any ex post bias must
go into slope estimates under the OLS estimation. So, second, the out-of-sample estimate of ERt
K
Interceptt ERM t k X kt
k 1
where ERMt = Rft + 0.05 is the estimated expected return for the market for the year and X k t is
the mean of explanatory variable in regression (2), Xkt, for all firms over the past ten years, the
period over which the mean βk are estimated. This calculation simply recognizes the property
that the OLS intercept is always the mean of the dependent variable minus the mean of the
explanatory variables multiplied by their estimated coefficients. But, rather than the mean Xkt
variable being that for the cross-section for the relevant year (which may be influenced by ex
post factors in that year), the mean, X k t , is now the mean over all firms for the preceding ten
years. The procedure recognizes that the mean expected return must be equal to the expected
return on the market. It does assume a risk premium of five percent, but the resultant cost-of-
capital estimate can be adjusted by an investor with a different required risk premium (price of
40
risk), and it can be adjusted for variation in that risk premium over time.37 As the revised
intercept just shifts ERt by a constant in the cross-section, the properties of the resultant cost-of-
Table 5 reports the mean cost-of-capital estimates, C of Ct, under the two procedures.
Under both procedures, cost-of-capital estimates around the median (portfolio 5) look close to
the standard expectation of 10 percent to 12 percent for the average return and close to the
average CAPM estimates in Table 4 (though benchmarking is difficult). They also align with
actual returns in t+1. The expected returns over portfolios co-vary strongly, with the Pearson
correlations ranging (for the mean X k t adjusted intercept) from 0.93 between portfolios 1 and 9 to
0.99 for a number of portfolios. This indicates that the estimates co-vary strongly with common
However, the estimates for the extreme portfolios are still extreme. These portfolios are
likely to be those with high measurement error (that throws estimates to the extremes). The
higher standard deviation of estimates for the extreme portfolios (not reported) suggest so. If an
estimate in the extreme is influenced by measurement error, subsequent estimates will regress
towards the error-corrected estimate provided that the measurement error is not strongly serially
correlated. So the table reports the median cost of capital for firms in each portfolio estimated
one year later, C of Ct+1. (Medians give lower weight to extreme observations in the extreme
portfolios.) For both adjustments, extremes are pulled closer to central values (though less so for
the risk-free rate intercept) and the estimates shrink over the whole range.
37
We applied an alternative procedure by predicting each Xkt from the following model estimated in time series
over the sample period: MeanXkt = a + b1.MeanXkt-1 + b2.MeanXkt-2 + b3.MeanXkt-3 + errort. Results were similar.
38
Rolling ten-year estimates follow the same pattern over time as those in Figure 2 (but with a mean shift).
41
The cost-of-capital estimates are those for one year ahead. The approach is amenable to
estimating regressions (1) and (2) with the target variables as growth and returns in subsequent
years. However, that could only be for the short-term future because of data limitations and
survivorship issues. Alternatively, one can envisage predictions from modeling the evolution of
the determining accounting variables in the future as investments are made and earnings are
deferred or realized. One can also envision a model that describes the evolution of the cost of
capital estimates over time, a model that then can be estimated and applied out of sample to
forecast the cost of capital recursively for a number of years ahead. Needless to say, forecasting
for the very long term is problematic, but investors presumably are primarily concerned with
shocks to their portfolios in the near-term, shocks that include revisions in expectations of the
6. A Simple Estimate
A reasonable criticism might point to the complexity of the estimation procedure. So we report
on a simpler, parsimonious estimate involving only Et/Pt and Bt/Pt. These are the two variables
that are identified in Penman, Reggiani, Richardson, and Tuna (2015) as jointly forecasting
earnings growth, the risk around the expected growth, earnings betas, and stock returns. They are
the variables which, when employed in the scheme in Exhibit 1, identify ROE and the effects of
conservative accounting on growth and risk in the Penman and Zhang (2015).
Et/Pt and Bt/Pt are the first two variables in regression equations (1) and (2). These
equations were estimated with just these two variables, and SERt (a simple expected return) was
then estimated by applying estimated coefficients from regression (2) out of sample, as before.
Table 6 summarizes the output for 10 portfolios formed on the estimate. The mean
intercept from estimating regression (2) over the years is 0.095 and the mean slope coefficients
42
on Et/Pt and Bt/Pt were 0.180 and 0.089, respectively.39 These estimates, applied to the mean out-
of-sample Et/Pt and Bt/Pt in the table, yield the SERt. The SERt are positively correlated with
actual t+1 returns. Cross-sectional regressions of actual returns for t+1 on SERt (like those in
Table 3) produced mean slope estimates of 1.135 for portfolios and 0.960 for individual firm
regressions, not significantly different from 1.0, and mean intercepts (average errors) not
significantly different from zero. The table also reports the shrinkage estimates for the cost of
capital with the intercept adjusted for the mean Xkt = (Et/Pt, Bt/Pt), as in Table 5.
This minimalist calculation uses just the bottom line numbers in the financial statements,
earnings, and book values. It impressive how well it performs. It is then a question of how much
improvement one gets by adding further information. The Aj variables in our analysis are such
information, but the question is open to further exploration, provided included variables accord
with our a priori criteria. Adding a specific measure of conservative accounting and its effect of
ROE may be helpful. Both operating and financing leverage might be incorporated. Penman,
Reggiani, Richardson, and Tuna (2015) show that the weights on Et/Pt and Bt/Pt in predicting
both earnings growth and returns changes with firm size, so the estimation might be done within
size groups. In a strict application of the scheme in Exhibit 1, one might estimate within Et/Pt
portfolios (where Et/Pt is roughly the same), effectively restricting the coefficient on Et/Pt in
regression (2) to be 1.0. As stated earlier, the point of this paper is to introduce an alternative
7. Conclusion
acceptance of any measure must be on the basis of what “looks good” relative to alternatives.
39
The mean estimates for 2001-2012 (which may be more relevant today) are 0.074 for the intercept and 0.112.and
0.104 for the coefficients on Et/Pt and Bt/Pt.
43
However, “looking good” requires a set of aesthetic criteria. We embrace the following. First, a
measure requires an a prioi rationale, and that rationale must involve a connection to risk.
Second that measure must be supported by empirical evidence that the a priori conditions are
satisfied. Third, in actual appearance, the measure must exhibit the features that one identifies
with risk and the expected return for risk. The measure here satisfies all three criteria, both in
absolute terms and on a relative basis to the alternatives examined. At worse, the analysis here is
though we view the paper in a much more positive and constructive light.
The cost of capital presumably changes over time as risk premiums (the price of risk)
change. So the estimates offered here are not necessarily those for a particular year. That would
depend on the price of risk at the time—indeed, each individual’s price of risk which presumably
varies over individuals. While Figure 1 indicates some common variation in expected return
estimates over time, the estimates are based largely on cross-sectional differences in risk rather
than the price of risk. An intriguing question is whether the latter can be extracted using
accounting data: Variation over time in mean Et/Pt and Bt/Pt or the other accounting
characteristics? The work of Ellahie, Katz, and Richardson (2014) that explains cross-country
average returns with aggregate Et/Pt and Bt/Pt is promising in this regard.
Appendix
44
Bt
This appendix lists the accounting variables, and Aj, j = 3,…, K, that enter as predictors of
Pt
earnings growth and returns in regressions (1) and (2). The variables are accompanied by an
explanation as to why they pertain to accounting principles that tie expected growth to risk and
thus are selected under requirement (i) in the text to then be examined under requirements (ii)
and (iii).
That explanation is accompanied by a report of whether requirements (ii) and (iii) are
supported empirically. For B/P, the sign of the mean estimated coefficient is that in the validating
regressions (1) and (2) with only E/P also in the regression. For the other variables, the sign is
from the validating regressions (1) and (2) with both E/P and B/P in the regression. Thus, the
sign of the mean coefficients for these variables are conditional upon the two bottom line
numbers, earnings and book values (and implicitly ROE), in the regression. See Penman and Zhu
A variable potentially indicates higher growth and risk if it indicates deferral of earnings
that results from the realization of expectations potentially indicates lower risk. While implied by
accounting principle, the a priori reasoning does not formally connect the variables to priced risk
and return. More modeling is required to make the explicit connection under a valid asset pricing
model (if one were available). However, where support is provided by extant asset pricing
theory, it is reported, for example with the connection of investment to expected returns in the
Merton inter-temporal asset pricing model or the Liu, Whited, and Zhang (2009) “Investment-
CAPM.”
45
Predicted and Predicted and
Calculation Actual Sign of Actual Sign of
Variable Explanation for Selection
Coefficient in Coefficient in
Regression (1) Regression (2)
Common equity at the end of fiscal-year t, Earnings and book value are the primary
divided by price at t. Book value is summary accounting numbers. For a given
Compustat’s common equity (item CEQ) E/P, B/P recovers earnings-to-book (ROE)
plus any preferred treasury stock (item that captures the effect of earnings deferral
TSTKP) less any preferred dividends in and conservative accounting on these Positive Positive
arrears (item DVPA). Book value and prices summary numbers. See text and Penman and
are on a per-share basis, with prices at three Zhang (2015) for further explanation and for
months after fiscal-year end adjusted for documentation that ties ROE to risky growth
Book-to-price
stock splits and stock dividends during the expectations and to returns. Penman,
(Bt/Pt)
three months after fiscal-year end. Reggiani, Richardson, and Tuna (2015)
document that, given E/P, B/P forecasts
earnings growth, and also the risk
surrounding the expected growth.
46
accelerated depreciation and amortization.
Change in gross property, plant, and Investment booked to the balance sheet is
equipment (item PPEGT) + change in the realization of (uncertain) investment
inventory (item INVT), all divided by lagged opportunities and thus resolution of the risk
assets that those opportunities will be realized. In
making investments, the firm signals that Negative Negative
earnings can be realized, and realized
earnings imply lower risk under accounting
principles. This is a restatement in
accounting terms of the connection between
the exercise of real options and expected
returns in Berk, Green, and Naik (1999). It is
also an elaboration on Cochrane (1991) q-
Investmentt
theory under which firms make investments
when the hurdle rate for investment is lower:
The prospect of realizable earnings lowers
that rate. Further, under conservative
accounting, investment booked to the
balances sheet is deemed lower risk for
earnings outcomes than investment that is
expensed immediately (as risky). The
Merton (1973) intertemporal CAPM and the
Liu, Whited, and Zhang (2009) Investment-
CAPM provide formal links to expected
returns.
Change in net operating assets divided by This is the total of all changes in the
average assets. Net operating assets is the operating section of the balance sheet due to
sum of accounts receivable (item RECT), realized income from operations and realized
inventory (item INVT), other current assets investments. It also includes recognized
(item ACO), property, plant, and equipment earnings not captured by the accruals
(item PPENB), intangible assets (item measure, for example, income in Negative Negative
Growth in net
INTAN) and other long term assets (item subsidiaries, realized gains and losses on
operating assets
AP), minus the sum of accounts payables assets sales, changes in deferred taxes,
(ΔNOAt)
(item AP), other current liabilities (item impairments and restructurings, and write-
LCO) and other long term liabilities (item downs. For a given earnings (in E/P), ΔNOA
LO) is lower if (relatively risky) investments are
expensed with conservative accounting.
Write-downs and impairments are due to
revisions of expected cash flows, and
47
expected cash flows are negatively related to
the cost of capital ceteris paribus in asset
pricing theory (see Johnstone 2015).
Change in debt plus the change in equity
from net equity transactions, scaled by External financing is positively correlated
average assets. Change in debt is the cash with current investment, and also indicates
proceeds from the issuance of long term debt plans for further investment. These
(item DLTIS) less cash payments for long investments are realizations (or anticipated
term debt reductions (item DLTR) plus the realizations) of uncertain investment
External financing
net changes in current debt (item DLCCH). opportunities which, in turn, signal that Negative Negative
(EXTFINt)
Change in equity is measured as the expected sales and earnings can be realized
proceeds from the sale of common and in the future (as above).
preferred stock (item SSTK) less cash
payments for the purchase of common and
preferred stock (item PRSKC) less cash
payments for dividends (item CDVC).
The natural log of the ratio of split-adjusted For given total financing (EXTFIN), net
shares outstanding at the end of the fiscal share issues reduce future earnings per share Negative Negative
Net share issue
year to shares outstanding at the end of the (growth) and reduce leverage.
(NSIt)
previous fiscal year.
48
References
Ang, A., J. Chen, and Y. Xing. 2006. Downside risk. Review of Financial Studies 19, 1191-1239.
Ashton, D., and P. Wang. 2013. Terminal Valuations, Growth Rates and the Implied Cost of
Capital. Review of Accounting Studies 18, 261-290.
Berk, J., R. Green, and V. Naik. 1999. Optimal investment, growth options, and security returns.
Journal of Finance 54, 1153-1608.
Blume, M. 1975. Betas and Their Regression Tendencies. Journal of Finance 30, 785-795.
Botosan, C., and M. Plumlee. 2005. Assessing Alternative Proxies for the Expected risk
Premium. The Accounting Review 80, 21-53.
Botosan., C., M. Plumlee, and J. Wen. 2011. The Relation Between Expected Returns, Realized
Returns, and Firm Risk Characterisitcs. Contemporary Accounting Research 28, 1085-1122.
Campbell, J. 1991. A Variance Decomposition for Stock Returns. Economic Journal 101, 157–
79.
Chattopadyhay, A., M. Lyle, and C. Wang. 2015. Accounting Data, Market Values, and the Cross
Section of Expected Returns Worldwide. Unpublished paper, Harvard University and
Northwestern University.
Chen, A. 2016. A General Equilibrium Model of the Value Premium with Time-Varying Risk
Premia. Unpublished paper, Federal Reserve Board, Washington, D.C.
http://ssrn.com/abstract=2763197
Christodoulou, D., C. Clubb, and S. McLeay. 2016. A Structural Accounting Framework for
Estimating the Expected Rate of Return on Equity. Forthcoming, Abacus 52, forthcoming.
Claus, J., and J. Thomas. 2001. Equity Risk Premium as Low as Three Percent? Evidence from
Analysts’ Earnings Forecasts for Domestic and International Stocks. Journal of Finance 56,
1629-1666.
Cochrane, J. 1991. Production-based asset pricing and the link between stock returns and
economic fluctuations. Journal of Finance 46, 209-237.
Daniel, K., and S. Titman. 1997. Evidence on the Characteristics of Cross-sectional Variation in
Common Stock Returns. Journal of Finance 52, 1-33.
Easton, P. 2004. PE Ratios, PEG Ratios, and Estimating the Implied Expected Rate of Return on
Equity Capital. The Accounting Review 79, 73-95.
49
Easton, P. 2007. Estimating the Cost of Capital Implied by Market Prices and Accounting Data.
Foundations and Trends in Accounting 2, 241-364, Now Publishers, Inc.
Easton, P., and S. Monahan. 2016. Review of Recent Research on Improving Earnings Forecasts
and Evaluating Accounting-based Estimates of the Expected Rate of Return on Equity Capital.
Abacus 52, forthcoming.
Easton. P., G. Taylor, P. Shroff, and T. Sougiannis. 2002. Using Forecasts of Earnings to
Simultaneously Estimate Growth and the Rate of Return on Equity Investment. Journal of
Accounting Research 40, 657-676.
Echterling, F., B. Eierle, and S. Ketterer. 2015. A Review of the Literature on Methods of
Computing the Implied Coast of Capital. International Review of Financial Analysis,
forthcoming.
Ellahie, A., M. Katz, and S. Richardson. 2014. Risky Value. Unpublished paper, London
Business School.
Fama, E., and K. French. 1993. Common risk factors in the returns of stocks and bonds. Journal
of Financial Economics 33, 3-56.
Fama, E., and K. French. 1997. Industry Cost of Capital. Journal of Financial Economics 43,
153-193.
Fama, E., and K. French. 2015a. A Five-Factor Asset Pricing Model. Journal of Financial
Economics 116, 1-22.
Fama, E., and K. French. 2015b. Dissecting Anomalies with a Five-Factor Model. Review of
Financial Studies 29, 69-103.
Feltham, G. and J. Ohlson. 1995. Valuation and clean surplus accounting for operating and
financial activities. Contemporary Accounting Research 11, 689-731.
Fitzgerald, T., Gray, S., Hall, J., and Jeyaraj, R. 2013. Unconstrained Estimates of the Equity
Risk Premium. Review of Accounting Studies 18, 560-639.
Gebhardt, W., C. Lee, and B. Swaminathan, 2001. Toward an Implied Cost of Capital.
Journal of Accounting Research 39, 135-176.
Gode, D., and P. Mohanram, 2003. Inferring the Cost of Capital Using the Ohlson-Juettner
50
Model. Review of Accounting Studies 8, 399-431.
Gode, D., and P. Mohanram. 2013. Removing Predictable Analysts Forecast Errors to Improve
Implied Cost of Equity Estimates. Review of Accounting Studies 18, 443-478.
Green, J., J. Hand, and F. Zhang. 2013. The Supraview of Return Predictive Signals. Review of
Accounting Studies 18, 692-730.
Green, J., J. Hand, and F. Zhang. 2014. The remarkable multidimensionality in the cross-section
of expected U.S. stock returns. Unpublished paper, University of North Carolina at Chapel Hill.
Guay, W., S. Kothari, and S. Shu. 2011. Properties of Implied Cost of Capital Using Analysts’
Forecasts. Australian Journal of Management 36, 125-149.
Harvey, C., Y. Liu, and H. Zhu. 2016. … and the cross-section of expected returns. Review of
Financial Studies 29, 5-68.
Hou, K., A. M. van Dijk, and Y. Zhang. 2012. The Implied Cost of Capital: A New Approach.
Journal of Accounting and Economics 53, 504-526.
Hou, K., C. Xue, and L. Zhang. 2015a. Digesting Anomalies: An Investment Approach. Review
of Financial Studies 28, 650-705.
Hou, K., C. Xue, and L. Zhang. 2015b. A Comparison of New Factor Models. Working paper,
Charles. A. Dice Center for Research in Financial Economics, Ohio State University.
Hughes, J., J. Liu, and J. Liu. 2009. On the Relation Between Expected Returns and Implied Cost
of Capital. Review of Accounting Studies 14, 246-259.
Johnstone. D. 2015. Information and the Cost of Capital in a Mean-Variance Efficient Market.
Journal of Business Finance and Accounting 42, 79-100.
Larocque, S. 2013. Analysts’ Earnings Forecast Errors and the Cost of Equity Capital Estimates.
Review of Accounting Studies 18, 135-166.
Lee, C., E. So, and C. Wang. 2015. Evaluating Firm-Level Expected-Return Proxies.
Unpublished paper, Stanford Graduate School of Business, MIT Sloan School of Management,
and Harvard Business.
Lewellen, J. 2015. The Cross-section of Expected Stock Returns. Critical Finance Review 4, 1-
44.
51
Li, K., and P. Mohanram. 2014. Evaluating Cross-sectional Forecasting Models for Implied Cost
of Capital. Review of Accounting Studies 19, 1152-1185.
Lyle, M., J. Callen, and R. Elliott. 2013. Dynamic Risk, Accounting-Based Valuation and Firm
Fundamentals. Review of Accounting Studies 18, 899-929.
Lyle, M. and C. Wang. 2015. The Cross Section of Expected Holding Period Returns and their
Dynamics: A Present Value Approach. Journal of Financial Economics 116, 505-525.
Miller, M., and F. Modigilani. 1961. Dividend Policy, Growth and the Valuation of Shares.
Journal of Business 34, 411-433.
Mohanram, P., and D. Gode. 2013. Removing Predictable Analyst Forecast Errors to Improve
Implied Cost of Capital Estimates. Review of Accounting Studies 18, 443-478.
Nekrasov, A., and M. Ogneva. 2011. Using Earnings Forecasts to Simultaneously Estimate Firm-
Specific Cost of Equity and Long-term Growth. Review of Accounting Studies 16, 414-457.
Ohlson, J. 1995. Earnings, Book Values, and Dividends in Equity Valuation. Contemporary
Accounting Research 12, 661-687.
Ohlson, J. 2008. Risk, Growth, and Permanent Earnings. Unpublished paper, Arizona State
University.
Penman, S. 2016. Valuation: Accounting for Risk and the Expected Return. Abacus 52,
forthcoming.
Penman, S., and F. Reggiani. 2013. Returns to buying earnings and book value: Accounting for
growth and risk. Review of Accounting Studies 18, 1021-1049.
Penman, S., and X. Zhang. 2015. Connecting Book Rate of Return to Risk and Return: The
Information Conveyed by Conservative Accounting. Unpublished paper, Columbia University
and University of California, Berkeley.
Penman, S., and J. Zhu. 2014. Accounting Anomalies, Risk and Return. The Accounting Review
89, 1835-1866.
Simin, T. 2008. The Poor Predictive Performance of Asset Pricing Models. Journal of Financial
and Quantitative Analysis 43, 355-380.
52
Stambaugh, R., and Y. Yuan. 2015. Mispricing Factors. Working paper No. 21533, National
Bureau of Economic Research.
Vasicek, O. 1973. A Note on Using Cross-sectional Information in Bayesian Estimation of
Security Betas. Journal of Finance 28, 1233-1239.
Zhang, L. 2015. The Investment CAPM. Unpublished paper, The Ohio State University.
Zhang, X. 2000. Conservative accounting and equity valuation. Journal of Accounting and
Economics 29, 125-149.
53
Figure 1. Ten-year Rolling Averages of Estimated Expected Returns, ERt, for Ten ERt Portfolios
0.350
0.300
Portfolio 10
0.250
0.200
0.150
0.100
0.050
Portfolio 1
0.000
-0.050
-0.100
54
Figure 2. Ten-year Rolling Averages of Predictive Slopes for Out-of-Sample ERt Forecasts of Actual Annual Returns
1.300
1.200
1.100
1.000
0.900
0.800
0.700
0.600
0.500
2000
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
55
Table 1
The table reports mean expected returns, ERt, and accounting characteristics for portfolios formed on the estimated expected returns. Reported numbers are
means over years of means for the ten portfolios formed each year, 1981-2012. Panel A characteristics are those at the time that ERt is estimated, and Panel B
characteristics are those observed subsequently. ERt is estimated by first estimating regression (2) in the cross-section with variables in the appendix, and then
applying mean coefficients estimates over the prior 10 years to accounting characteristics in each year on a rolling basis. Forecasts of EPS growth rates two years
ahead are estimated by applying mean coefficients from estimating regression equation (1) over the prior 10 years to accounting characteristics in each year.
56
Panel B: Accounting Characteristics Subsequent to Time-t for Expected Return Portfolios
Expected
Return
Portfolio EPS Growth Rate Two Years Ahead EPS Over Next Four Years/Pt
Earnings
STD IDR STD IDR Earnings Change
Forecast Actual Actual Actual Actual Actual Actual Betat+1 Betat+1
1 (Low) 0.119 0.023 0.154 0.361 0.013 0.072 0.192 0.45 0.84
2 0.043 0.008 0.157 0.459 0.091 0.048 0.112 0.71 0.81
3 0.030 0.012 0.163 0.425 0.134 0.050 0.098 0.72 0.94
4 0.028 0.019 0.147 0.382 0.154 0.049 0.115 0.85 0.69
5 0.031 0.044 0.147 0.430 0.184 0.052 0.109 0.85 0.47
6 0.041 0031 0.176 0.462 0.188 0.052 0.066 0.90 0.72
7 0.053 0.053 0.168 0.450 0.206 0.054 0.091 1.16 1.27
8 0.070 0.051 0.174 0.376 0.203 0.060 0.138 1.30 1.82
9 0.104 0.082 0.206 0.502 0.200 0.081 0.218 2.10 2.24
10 (High) 0.215 0.123 0.263 0.672 0.162 0.124 0.308 4.42 2.46
Et/Pt is earnings for fiscal-year t divided by stock price. Earnings are before extraordinary items (Compustat item IB) and special items (item SPI), minus
preferred dividends (item DVP), with a tax allocation to special items at the prevailing Federal statutory corporate income tax rate for the year. Earnings and
prices are on a per-share basis, with prices observed three months after fiscal-year end adjusted for stock splits and stock dividends during the three months after
fiscal year end. Operating profit margin is earnings before net interest expense divided by sales. All other accounting characteristics in Panel A are defined in the
appendix.
In Panel B, STD is standard deviation and IDR is the interdecile range of annual portfolio means over years. EPS Over Next Four Years/Pt is the sum of split-
adjusted EPS for years t+1 to t+4 with dividends for year t+1 to t+3 reinvested as the prevailing risk-free rate for the year, all divided by price per share at t.
Earnings beta is the slope coefficient from estimating the following time-series regression of portfolio annual earnings yield on the market-wide earnings yield,
for December 31 fiscal-year firms only (to align in calendar time):
57
Earningst 1 Earningst 1
Portfolio Market t . The portfolio earnings yield is aggregate earnings for the portfolio relative to aggregate price and the
Pt Pt
market earnings yield is aggregate earnings for all stocks in the sample for the relevant year relative to aggregate price. Earnings change betas are similarly
estimated. ERt+1 is the mean expected return at the end of year t+1 estimated for firms in the respective ER t portfolios at time t.
58
Table 2
Expected Returns, Actual Returns, and Forward Betas
The table summarizes the distribution of actual (realized) annual returns for ERt portfolios over the following year. It also reports betas for these actual returns
with respect to the market return (the return for the value-weighted CRSP index) and their sensitivities to factors in the Fama and French (FF) five-factor model
and I/A factor in the Hou, Xue, and Zhang (2015a) (HXZ) Investment CAPM.
All Up Down
Mean Median STD Range Skewness Kurtosis HML SMB RMW CMA I/A
Years Beta Beta
1 (Low) 0.043 0.049 0.028 0.258 0.977 0.495 2.364 1.31 1.77 0.48 0.07 0.84 -0.31 -0.34 -0.46
2 0.117 0.094 0.042 0.211 0.847 0.074 2.344 1.13 1.35 0.81 0.10 0.78 -0.15 -0.15 -0.29
3 0.145 0.129 0.117 0.216 0.967 0.097 2.886 1.15 1.44 0.87 0.09 0.78 -0.13 -0.08 -0.15
4 0.162 0.144 0.139 0.214 1.037 0.557 3.562 1.12 1.38 0.85 0.06 0.70 -0.11 0.01 -0.04
5 0.174 0.155 0.113 0.215 1.041 0.497 3.330 1.20 1.58 0.96 0.07 0.75 -0.03 0.06 0.02
6 0.186 0.160 0.151 0.197 0.941 0.485 3.784 1.18 1.62 0.99 0.19 0.73 -0.05 0.02 0.10
7 0.197 0.179 0.174 0.201 0.990 0.337 3.569 1.20 1.65 1.12 0.17 0.70 -0.09 0.06 0.13
8 0.209 0.191 0.174 0.235 1.074 0.717 4.316 1.30 1.77 1.27 0.19 0.72 -0.14 0.04 0.22
9 0.225 0.212 0.219 0.258 1.413 1.142 6.050 1.41 2.02 1.44 0.27 0.73 -0.32 0.01 0.26
10 (High) 0.264 0.282 0.241 0.360 1.826 1.265 5.782 1.82 2.89 1.45 0.14 0.80 -0.62 0.13 0.28
Actual return metrics summarize the times series of portfolio actual returns for ERt portfolios in year t+1, 1981-2012. Betas and other factor betas are estimated
from the time series of portfolios returns, the market return betas with annual returns and the Fama and French (FF) betas with monthly returns. Returns are
equally weighted in ERt portfolios. Market return betas are estimated with December 31 fiscal-year firms only. Up markets are those where the CRSP value-
weighted index was greater than 10% for the year, and down markets are those where it was less than -10%. FF and HXZ betas are estimated with all factors in
59
the model (not one at a time). HML, SMB, RMW, and CMA monthly returns are those for the Fama and French book-to-price, size, profitability, and investment
factors, respectively, and are from the Kenneth French website library. Monthly returns on Investment CAPM factors were supplied by Lu Zhang. With the
exception of portfolio1, none of the estimated betas on the CMA factor are significantly different from zero, nor are those for HML in portfolios 1-5 and RMW in
portfolios 5 and 6. The t-statistics on the I/A factor are significantly different from zero, with the exception of portfolios 4 and 5. Betas on other Investment
CAPM factors (size and ROE) were similar across portfolios to the corresponding factors in the FF model.
60
Table 3
Coefficient Estimates from Annual Cross-sectional Regressions of Forward Actual Returns
(Rt+1) on ERt and Other Variables
The table reports mean coefficients from estimating cross-sectional regressions each year, 1981-2012. The t-
statistics (in parenthesis) are estimated from the time-series of estimated coefficients with a Newey-West correction for
the serial correlation in the coefficient estimates. Size is log market capitalization of equity. Beta is the return beta on the
value-weighted CRSP index return. Leverage is net debt/market value of equity. Momentum is the buy-and-hold
return over the twelve months prior to one month before the time t when ERt is estimated.
61
Table 4
Implied Cost-of-Capital Estimates (ICC) and Cost-of-Capital Estimates from the Capital Asset Pricing Model (CAPM) and
Fama and French Three-Factor Model (FF)
For each ERt portfolio, Panel A of the table reports mean ICC estimates calculated as in Gebhardt, Lee, and Swaminathan (2001) (GLS) and Easton, Taylor,
Shroff, and Sougiannis (2002) (ETSS), with mean intercept and slope coefficients and implied cost of capital, r, and implied growth, g, from applying procedures
in ETSS. Panel B reports the mean cost of capital, C of Ct, and actual returns for portfolios formed from CAPM estimates and Fama and French (FF) model
estimates with three factors (MKT, SMB, and HML), along with related measures. It also reports the mean CAPM and FF cost-of-capital estimates for ERt
portfolios and, for the latter, the intercepts and associated t-statistics from time-series regressions of excess monthly returns for the ERt portfolios.
62
Panel B: Cost of Capital Estimates from Asset Pricing Models (Full Sample Factor Premium)
1 0.069 0.06 0.75 0.142 0.133 1.36 0.006 0.128 0.161 -0.0065 -3.92
2 0.091 0.50 0.80 0.159 0.129 1.27 0.071 0.148 0.155 -0.0026 -2.23
3 0.101 0.71 0.90 0.170 0.127 1.22 0.096 0.157 0.151 -0.0003 -0.28
4 0.109 0.88 0.97 0.170 0.124 1.17 0.116 0.167 0.149 0.0002 0.24
5 0.117 1.03 1.06 0.170 0.123 1.15 0.135 0.160 0.147 0.0019 2.27
6 0.125 1.19 1.15 0.168 0.122 1.14 0.153 0.159 0.149 0.0017 2.11
7 0.134 1.38 1.21 0.161 0.121 1.12 0.174 0.159 0.149 0.0034 4.00
8 0.146 1.61 1.31 0.161 0.121 1.11 0.200 0.168 0.149 0.0038 4.17
9 0.163 1.94 1.46 0.135 0.121 1.11 0.238 0.161 0.152 0.0049 4.32
10 0.202 2.73 1.67 0.130 0.120 1.09 0.358 0.158 0.154 0.0081 4.99
Historical CAPM Betat is estimated for each firm from monthly returns over 60 months prior to the point when the Cof Ct is estimated
(or a lesser period if returns are not available for the full period). Betat+1 is estimated from monthly portfolio returns during actual
return periods over the complete sample period. Fama and French cost-of-capital estimates are calculated by applying factor
sensitivity coefficients estimated over the prior 60 months to mean factor returns over the whole sample and adding the estimated
intercept, with these estimates then annualized. The FF intercepts are estimated from FF model time-series regressions with monthly
excess returns for ERt portfolios during actual returns periods.
63
Table 5
The table reports cost-of-capital estimates, C of Ct, for portfolios, calculated after shrinking the expected returns in Tables 1 and 2 with an intercept adjustment
when applying regression (2). The first adjustment sets the intercept at the prevailing risk-free rate for the year. The second adjusts the intercept for the mean of
the explanatory variables over the past ten years. The table also reports mean actual returns in the subsequent year and the mean estimated cost of capital at the
end of that year, C of Ct+1, for firms in the respective C of Ct portfolios.
64
Table 6
The table reports means (over years) of Simple Expected Returns, SERt, actual returns, and shrinkage cost-of-capital estimates for portfolios formed on SERt
each year, 1981-2012. SERt is estimated out-of-sample by applying mean regression coefficients over the past ten years from estimating regression (2) with only
Et/Pt and Bt/Pt as explanatory variables. The table also reports mean regression coefficients from estimating regression (2) in all years, 1981-2012. The shrinkage
estimates are mean-adjusted estimates, as in the second adjustment in Table 5.
65