This action might not be possible to undo. Are you sure you want to continue?
https://www.scribd.com/doc/77546191/2FrankelLeeJAE1998
02/05/2015
text
original
¹ Examples of Ohlson’s work include Ohlson (1990, 1991, 1995) and Feltham and Ohlson (1995).
Examples of empirical research include Bernard (1994), Fairﬁeld (1994), Ou and Penman (1994),
Penman and Sougiannas (1996), Abarbanell and Bernard (1995), and Frankel and Lee (1998), Lee
et al. (1998), and Dechow et al. (1997).
Journal of Accounting and Economics 25 (1998) 283—319
Accounting valuation, market expectation,
and crosssectional stock returns
Richard Frankel, Charles M.C. Lee*
School of Business Administration, University of Michigan, Ann Arbor, MI 481091234, USA
Johnson Graduate School of Management, Cornell University, Ithaca, NY 148534201, USA
Received 1 May 1997; accepted 7 August 1998
Abstract
This study examines the usefulness of an analystbased valuation model in predicting
crosssectional stock returns. We estimate ﬁrms’ fundamental values (») using I/B/E/S
consensus forecasts and a residual income model. We ﬁnd that »is highly correlated with
contemporaneous stock price, and that the »/P ratio is a good predictor of longterm
crosssectional returns. This eﬀect is not explained by a ﬁrm’s market beta, B/P ratio, or
total market capitalization. In addition, we ﬁnd errors in consensus analyst earnings
forecasts are predictable, and that the predictive power of »/P can be improved by
incorporating these errors. 1998 Elsevier Science B.V. All rights reserved.
JEL classiﬁcation: D4; G12; G14; M4
Keywords: Capital markets; Market expectations; Market eﬃciency; Valuation; Analyst
forecasts
1. Introduction
Recent studies by Ohlson on residual income valuation have led empirical
researchers to reexamine the relation between accounting numbers and ﬁrm
value.¹ In this study, we operationalize the residual income model using analyst
01654101/98/$ — see front matter 1998 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 5  4 1 0 1 ( 9 8 ) 0 0 0 2 6  3
` Several studies show analyst forecast errors diﬀer for ﬁrms with certain characteristics, sugges
ting a relation between analyst forecast errors and various market pricing anomalies (e.g., Dechow
and Sloan, 1997; Daniel and Mande, 1994; LaPorta, 1996; LaPorta et al., 1997). Other studies show
earnings forecasts and examine its usefulness in predicting crosssectional stock
returns in the U.S. Speciﬁcally, we use I/B/E/S consensus earnings forecasts to
proxy for market expectations of future earnings. We then use the resulting
estimate of ﬁrm fundamental value (»
) to investigate issues related to market
eﬃciency and the predictability of crosssectional stock returns.
We ﬁnd that »
estimates based on I/B/E/S consensus forecasts are highly
correlated with contemporaneous stock prices. In recent years, »
explains more
than 70% of the crosssectional variation in stock prices. Moreover, the value
toprice ratio (»
/P) is a good predictor of crosssectional returns, particularly
over longer time horizons. In 12month horizons, the »
/P ratio predicts
crosssectional returns as well as the booktomarket ratio (B/P). However, over
two or three year periods, buyandhold returns from »
/P strategies are more
than twice those from B/P strategies. Speciﬁcally, we ﬁnd that higher »
/P ﬁrms
tend to earn higher longterm returns. This result is not due to diﬀerences in
market betas, ﬁrm size, or the B/P ratio.
Because of its importance in estimating »
, we also investigate the reliability
of longterm I/B/E/S consensus earnings forecasts. We ﬁnd that crosssectional
errors in the threeyearahead consensus forecast are predictable. Speciﬁcally,
we ﬁnd some evidence that analysts tend to be more overlyoptimistic in ﬁrms
with higher past sales growth (SG) and higher P/B ratios. In addition, we ﬁnd
stronger evidence of overoptimism in ﬁrms with higher forecasted earnings
growth (¸tg) and higher forecasted ROEs relative to current ROEs (OP).
Combining these variables in a prediction model, we develop an estimate of the
prediction error in longterm forecasts (PErr), and show this estimate has
predictive power for crosssectional returns.
Finally, we show the predictive power of PErr is incremental to a »
/P
strategy. During our sample period (1979—1991), a zerocash investment strategy
involving ﬁrms that are simultaneously in the top quintile of »
/P and the
bottom quintile of PErr yields cumulative buyandhold returns of more than
45% over 36 months. The threeyear buyandhold strategy results in positive
returns in both up and down markets. This eﬀect is not explained by market
beta, ﬁrm size, or the B/P ratio.
Our results contribute to the emerging literature on the residual income
model in several ways. First, our analystbased approach complements Penman
and Sougiannas (1998), which uses ex post reported earnings. Second, we
provide evidence on the reliability of I/B/E/S consensus forecasts for valuation,
as well as a method for correcting predictable forecast errors. To our knowledge,
this is the ﬁrst study to develop a prediction model for longrun analyst forecast
errors, and to trade proﬁtably on that prediction.` Finally, we show that returns
284 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
analysts may not use all available information when formulating their forecasts (e.g., Abarbanell,
1991; Abarbanell and Bernard, 1992; Stober, 1992). However, none of these studies develop
a prediction model for analyst errors. Brown et al. (1995) do develop a prediction model for analyst
errors, but their investment horizon is only onequarterahead and the details of their model are
proprietary.
` The term Edwards—Bell—Ohlson, or EBO, was coined by Bernard (1994). Theoretical develop
ment of this valuation method is found in Ohlson (1990, 1995), Lehman (1993), and Feltham and
Ohlson (1995). Earlier treatments can be found in Preinreich (1938), Edwards and Bell (1961), and
Peasnell (1982). For a simple guide to implementing this technique, see Lee (1996).
to a »
/P strategy are not due to standard risk proxies, and that the strategy can
be further improved by incorporating analyst forecast errors.
Our ﬁndings are also related to the ﬁnance literature on the predictability of
stock returns. Much recent research has focused on accountingbased ratios that
exhibit predictive power for stock returns. The B/P ratio, in particular, has been
elevated to celebrity status by studies such as Fama and French (1992). Fama
and French suggest B/P is a proxy for a ﬁrm’s distress risk. However, little
progress has been made in identifying the exact nature of this risk. Our results
suggest that rather than attempting to produce a better risk proxy, superior
return prediction may result from adopting a more complete valuation ap
proach.
In sum, empirical studies involving equity valuation encounter two potential
problems: (1) the use of overly restrictive models of intrinsic value, and (2) the
use of biased proxies as model imputs. Our research design features a more
robust valuation model than simple marketmultiples, as well as a technique for
improving on analysts’ earnings forecasts. Our empirical ﬁndings suggest both
will lead to better predictions of crosssectional stock returns.
The remainder of this paper is organized as follows. In the next section, we
present the accountingbased valuation model and describe its most salient
features. In Section 3, we discuss the estimation procedures used to implement
this model. Section 4 contains a discussion of the data and sample description.
Section 5 reports the empirical results, and Section 6 concludes with a summary
of our ﬁndings and their implications.
2. The residual income model
The valuation method we use in this study is a discounted residual income
approach sometimes referred to as the Edwards—Bell—Ohlson (EBO) valuation
technique.` Independent derivations of this valuation model have surfaced
periodically throughout the accounting, ﬁnance and economics literature since
the 1930s. In this section, we present the basic residual income equation and
brieﬂy develop the intuition behind the model.
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 285
A stock’s fundamental value is typically deﬁned as the present value of its
expected future dividends based on all currently available information. Nota
tionally,
»H
R
,
`
G¯¹
E
R
(D
R>G
)
(1#r
)G
. (1)
In this deﬁnition, »H
R
is the stock’s fundamental value at time t, E
R
(D
R>G
) is the
expected future dividends for period t#i conditional on information available
at time t, and r
is the cost of equity capital based on the information set at time
t. This deﬁnition assumes a ﬂat termstructure of discount rates.
It is easy to show that, as long as a ﬁrm’s earnings and book value are
forecasted in a manner consistent with clean surplus accounting, Eq. (1) can be
rewritten as the reported book value, plus an inﬁnite sum of discounted residual
income:
»H
R
"B
R
#
`
G¯¹
E
R
[NI
R>G
!(r
B
R>G¹
)]
(1#r
)G
"B
R
#
`
G¯¹
E
R
[(ROE
R>G
!r
) B
R>G¹
]
(1#r
)G
, (2)
where B
R
is the book value at time t, E
R
[.] is expectation based on information
available at time t, NI
R>G
is the Net Income for period t#i, r
is the cost of
equity capital and ROE
R>G
is the aftertax return on book equity for period t#i.
Note that this equation is identical to a dividend discount model, but
expresses ﬁrm value in terms of accounting numbers. It therefore relies on the
same theory and is subject to the same theoretical limitations as the dividend
discount model. However, the model provides a framework for analyzing the
relation between accounting numbers and ﬁrm value.
Eq. (2) shows that equity value can be split into two components — an
accounting measure of the capital invested (B
R
), and a measure of the present
value of future residual income, deﬁned as present value of future discounted
cash ﬂows not captured by the current book value. If a ﬁrm earns future
accounting income at a rate exactly equal to its cost of equity capital, then the
present value of future residual income is zero, and »
R
"B
R
. In other words,
ﬁrms that neither create nor destroy wealth relative to their accountingbased
shareholders’ equity, will be worth only their current book value. However,
ﬁrms whose expected ROEs are higher (lower) than r
will have values greater
(lesser) than their book values.
If the market price approximates future discounted cash ﬂows, then Eq. (2)
oﬀers a natural interpretation for the pricetobook ratio. Dividing both sides
of Eq. (2) by B
R
, we can express P/B in terms of a ﬁrm’s future abnormal
ROEs. In a competitive equilibrium, a typical ﬁrm’s ROE should be close to its
286 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
" In practice, ﬁrms’ reported ROE may diﬀer from their costs of equity in competitive equilibrium
due to accounting and risk factors. For example, we show later (in Table 1) that average P/Bs are
above 1, and average ROEs are somewhat above ﬁrms’ costs of capital. These observations are
consistent with the fact that accounting systems that are, on average, conservative.
cost of equity capital (ROE"r
), and the typical pricetobook ratio should be
close to 1." Moreover, ﬁrms expected to earn above (below) normal ROEs in the
future should trade at higher (lower) pricetobook ratios. Stated another way,
this equation properly impounds a ﬁrm’s expected future proﬁtability into its
equity value estimate.
Eq. (2) shows that future earnings performance should be closely linked to
current B/P ratios. This association is consistent with the empirical ﬁnding
that low (high) B/P ﬁrms have higher (lower) future ROEs (e.g., FF, 1995;
Fairﬁeld, 1994; Bernard, 1994). However, the inverse relation between future
earnings performance and current B/P ratios should not be interpreted as
proof of market eﬃciency — it shows that the market considers future proﬁtabil
ity when formulating prices, not that it fully incorporates all available informa
tion when doing so. Later, we use analyst forecasts of future earnings to directly
examine whether the market fully incorporates current information in establish
ing prices.
Current literature shows that a number of ad hoc variables such as cash ﬂow
yield (Chan et al., 1991; LSV, 1994; Davis, 1994), earnings yield (Basu, 1977;
Jaﬀe et al., 1989) and dividend yield (Litzenberger and Ramaswamy, 1979) have
predictive power for crosssectional returns. These yield measures have
often been interpreted as risk proxies. Eq. (2) suggests that these market
multiples may also work because their accounting component reﬂects (imper
fectly) some dimension of »H
R
. However, neither book value nor earnings, is
suﬃcient to capture »H
R
. One of the objectives in this paper is to use the
EBO model to derive a more precise estimate of »H
R
, and examine whether
a more complete valuation model yields superior power to predict risk adjusted
returns.
Several recent studies evaluate this model’s ability to explain stock prices.
Penman and Sougiannas (1998) implement variations of the model using ex post
realizations of earnings to proxy for ex ante expectations. Lee et al. (1998)
operationalize the model for the 30 stocks in the Dow Jones Industrial Average
and examine timeseries properties of the model. Frankel and Lee (1998) employ
the model in an international context and ﬁnd that » has high explanatory
power for prices in 21 countries. More recently, both Francis et al. (1997) and
Dechow et al. (1997) examine the empirical properties of the model under
alternative speciﬁcations. Except for Dechow et al. (1997), these studies do not
examine the predictive power of the model for crosssectional stock returns in
the US.
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 287
` Six percent reﬂects the average longrun returnonassets (see Table 1, last column). We use this
measure as a proxy for normal earnings when reported earnings are negative. As explained later, we
also constrain k to be between 0 and 100%.
3. Model estimation procedures
Eq. (2) presents a simple procedure for estimating a ﬁrm’s intrinsic value (»
R
).
The four parameters needed for the estimation are: the cost of equity capital (r
),
future ROE forecasts (FROEs), current book value (B
R
), and a dividend payout
ratio (k). The ﬁrst three parameters’ roles are readily seen in Eq. (2). The last
input, the dividend payout ratio (k), is used in conjunction with the clean surplus
relation (CSR) to derive future book values. In this section, we discuss the
speciﬁcs of the model estimation procedure.
Cost of equity capital (r
). In theory, r
should be ﬁrmspeciﬁc, reﬂecting the
premium demanded by equity investors to invest in a ﬁrm or project of
comparable risk. In practice, however, there is little consensus on how this
discount rate should be determined. For this study, we use three diﬀerent
approaches — a constant discount rate, and two industrybased discount rates
derived by FF (1997). The FF discount rates are based on a onefactor and
a threefactor risk model.
Dividend payout ratio (k). The dividend payout ratio is the percentage of net
income paid out in the form of dividends each year. We obtain a ﬁrmspeciﬁc
estimate of k by dividing the common stock dividends paid in the most recent
year (Compustat Item 21) by net income before extraordinary items (Compustat
Item 237). For ﬁrms with negative earnings (approximately 11% of our sample),
we divide dividends by six percent of total assets to derive an estimated payout
ratio.` This variable is used, in conjunction B
R
, to derive forecasted book values:
B
R>¹
"B
R
#NI
R>¹
!d
R>¹
"B
R
#(1!k)NI
R>¹
"[1#(1!k)ROE
R>¹
]B
R
.
Analogously, all future book values can be expressed as functions of B
R
, k, and
future ROEs. For example, we can write
B
R>`
"[1#(1!k)ROE
R>¹
][1#(1!k)ROE
R>`
]B
R
.
Future ROEs. The most important and diﬃcult task in the EBO valuation
exercise is forecasting future ROEs (or, equivalently, forecasting future earn
ings). Two alternatives, based on ex ante information, are: (1) use prior period
earnings (or ROEs), or (2) use analysts’ earnings forecasts (e.g., Abarbanell and
Bernard (1995) use earnings forecasts from ValueLine]. We use both methods
and derive a value metric based on historical earnings (»
), as well as a value
metric based on consensus I/B/E/S analyst forecasts (»
).
288 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
" Abarbanell and Bernard (1995) use ValueLine forecasts to estimate a similar EBO valuation
equation in addressing the question of whether U.S. markets are myopic. Under the null hypothesis
of market eﬃciency, they examine whether markets underprice longrun earnings relative to
nearterm earnings. They also provide evidence that future forecasted ROEs may not be fully
impounded in current prices.
` We also estimated a 12period expansion of the formula in which ROEs are reverted back to the
industry median. The 12period version had slightly lower correlation with stock prices and similar
predictive power for returns.
Fairﬁeld et al. (1994) show that, in large samples, the correlation between
current year ROEs and next year’s ROEs is around 0.66, suggesting that the
current period ROE is a reasonable starting point for estimating future ROEs.
The use of I/B/E/S data should result in a more precise proxy for market
expectations of earnings. Prior studies show that analyst earnings forecasts are
superior to timeseries forecasts (e.g., O’Brien, 1988; Brown et al., 1987a,b).
However, the predictive superiority of an analystbased value metric (»
) over
a historicalbased value metric (»
) is an open empirical question."
Forecast horizons and terminal value estimation. Eq. (2) expresses ﬁrm value in
terms of an inﬁnite series, but for practical purposes, the explicit forecast period
must be ﬁnite. This limitation necessitates a terminal value estimate — that is, an
estimate of the value of the ﬁrm based on residual income earned after the
explicit forecasting period. One approach is to estimate the terminal value by
ﬁrst expanding Eq. (2) to ¹ terms, and then taking the next term in the
expansion as a perpetuity. For example, if the explicit forecast period ends after
¹ periods, the terminal value is:
(ROE
2>¹
!r
)
(1#r
)2r
B
2
.
This procedure is mathematically equivalent to a ¹period discounted dividend
model in which year ¹#1 earnings is treated as a perpetuity (see Penman,
1995). The resulting value estimate therefore depends critically on the particular
earnings forecast used in the terminal value. Various alternative approaches
have appeared in the literature. For example, both Lee et al. (1998) and Dechow
et al. (1997) feature various permutations for the terminal value.
In this study, we take a simple approach using a shorthorizon earnings
forecasts of up to three years.` In theory, ¹ should be set large enough for ﬁrms
to reach their competitive equilibrium. However, our ability to forecast future
ROEs diminishes quickly over time, and forecasting errors are compounded in
longer expansions. Therefore, we estimate three forms of »
R
:
»K¹
R
"B
R
#
(FROE
R
!r
)
(1#r
)
B
R
#
(FROE
R
!r
)
(1#r
)r
B
R
, (3.1)
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 289
` In theory, the model calls for beginningofyear book values. However, we use the annual
average to avoid situations where an unusually low book value in year t1 inﬂates forecasted ROEs.
»K`
R
"B
R
#
(FROE
R
!r
)
(1#r
)
B
R
#
(FROE
R>¹
!r
)
(1#r
)r
B
R>¹
, (3.2)
»K`
R
"B
R
#
(FROE
R
!r
)
(1#r
)
B
R
#
(FROE
R>¹
!r
)
(1#r
)`
B
R>¹
#
(FROE
R>`
!r
)
(1#r
)`r
B
R>`
. (3.3)
Eq. (3.1) represents a twoperiod expansion of the residual income model with
the forecasted ROE for the current year (FROE
R
) assumed to be earned in
perpetuity. Eq. (3.2) also represents a twoperiod expansion of the model, but we
use a twoyearahead forecasted ROE (FROE
R>¹
)in the perpetuity. Similarly,
Eq. (3.3) is a threeperiod model.
The righthand side of each equation consists of ex ante observables. To
estimate »
, we use the return on average equity, ROE
R
"NI
R
/[(B
R
#B
R¹
)/2],
to proxy for all future ROEs — i.e., we substitute ROE
R
for all the FROEs in the
above equations.` NI
R
is earnings to common shareholders in year t, net of
extraordinary items, taxes, and preferred dividends (Compustat Item 237), and
B
R
is total common shareholders’ equity from year t (Compustat Item 60). To
estimate »
, we derive future ROEs and book values from I/B/E/S consensus
forecasts using a sequential procedure described in the Appendix.
4. Data and sample description
The original sample of ﬁrms consists of all domestic nonﬁnancial companies
in the intersection of (a) the NYSE, AMEX, and NASDAQ return ﬁles from the
Center for Research in Security Prices (CRSP) and (b) a merged Compustat
annual industrial ﬁle, including PST, full coverage and research ﬁles. We require
ﬁrms to meet the Compustat data requirements (for B
R¹
, B
R`
, NI
R¹
, and
DI»
R¹
) and have the necessary CRSP stock prices and shares outstanding data
(for ﬁscal year end t!1, and the end of June in year t). Furthermore, we require
ﬁrms to have a oneyearahead and a twoyearsahead earningspershare (EPS)
forecast from I/B/E/S. Because I/B/E/S began operations in 1975, this require
ment limits our sample period to 1975—93. We further constrain our sample to
ﬁrms with ﬁscalyearends between June and December, inclusively. Because we
use I/B/E/S forecasts issued in May, this constraint ensures that forecasted
earnings correspond to the correct ﬁscal year. Using »
for the entire Fama and
French (1992) sample over the period of 1962—1993 yields similar results. The
results are also similar for a sample consisting of just December yearend ﬁrms.
290 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
" Using median rather than mean forecasts is unlikely to aﬀect results because the distribution of
forecasted growth is quite symmetric.
¹" See Ball et al. (1995) for additional evidence on the sensitivity of contrarian returns to the
removal of stocks with prices under $1.
To ensure that accounting variables are known before returns are computed,
we allow a minimum gap of six months between the ﬁscalyearend and the
portfolio formation date. Speciﬁcally, we match accounting data for all ﬁscal
year ends in the calendar year t!1 to returns on portfolios formed at the
end of June of year t. We use a ﬁrm’s market equity at its ﬁscalyearend to
compute its booktomarket and valuetomarket ratios, and the market
equity on June 30 of year t to measure its size. These procedures are similar to
FF (1992), except their B/P ratios are based on December market prices, rather
than ﬁscal year end. They report that this diﬀerence has little eﬀect on their
return tests.
In estimating »
, we use the I/B/E/S mean (also called consensus) forecast
from the May statistical period of year t. This mean estimate is determined from
analyst forecasts on ﬁle with I/B/E/S as of the Thursday after the third Friday of
each month." Since these monthly reports are widely available soon after each
computer run, the May statistics are in the public domain well before our
portfolio formation date. Our valuation formula uses three pieces of I/B/E/S
data: earningspershare forecasts oneyearahead (F½1), EPS forecasts two
yearsahead (F½2), and a ﬁveyear longterm growth rate (¸tg). Between 1975
and 1979, analysts reported just F½1 and F½2, but after 1980, most ﬁrms also
had ¸tg information. The Appendix explains the procedure we followed to
derive future ROE forecasts when all three variables are not in the May I/B/E/S
report.
In estimating Eqs. (3.1), (3.2) and (3.3), we remove ﬁrms with negative book
values, because ROEs for these ﬁrms cannot be interpreted in economic terms.
In addition, some ﬁrms have extremely low book values, or earnings, leading to
unreasonable ROE or k estimates. We eliminate such ﬁrms by considering only
ﬁrms with ROEs or FROEs of less than 100% and dividend payout ratios of less
than 100%. These procedures eliminate 1075 ﬁrmyears. We also remove 51
ﬁrms with stock prices of under $1 as of the end of June in year t. These ﬁrms
have unstable B/P, »
/P and »
/P ratios and poor market liquidity (that is, they
cannot be included in equalweighted portfolios without incurring dispropor
tionally large trading costs).¹"
Taken together, our ﬁlters eliminated 1,126 observations (approximately 5%),
leaving a ﬁnal sample of 18,162 ﬁrmyears. These common sense ﬁlters ensure
the subsequent results are not driven by outliers. Further, the strategies we
examine are tradable, in the sense that all the portfolios are constructed using
ﬁrm characteristics that are observable at the time of portfolio formation.
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 291
Table 1
Summary statistics by year
Table values represent annual, equallyweighted average statistics for the sample ﬁrms. Year t!1 is
the year fromwhich the accounting data are obtained. ME is the market value of equity as of June 30
of year t in millions of dollars. k is the dividend payout ratio, computed as common stock dividends
divided by earnings to common shareholders. For ﬁrms with negative earnings, k is computed as
common stock dividends divided by (total assets ;0.06). ROE is the return on equity for year t!1
computed as net income for year t!1 divided by the year t!1 average book equity. B is the year
t!1 reported book value per share. P is the stock price as of June 30 in year t. ROA is the return on
total assets for year t!1. 1/(Avg.B/P) is the inverse of the equallyweighted average B/P ratio for
each year. Averages reported in the bottom row represent timeseries means of the annual statistics.
Year t No.
ﬁrm
Avg.
ME
Avg.
k
Avg.
ROE
Avg.
B
Avg.
P/B
1/(Avg.
B/P)
Avg.
ROA
76 361 1168 0.35 0.14 21.77 2.02 1.40 0.07
77 312 1264 0.32 0.16 22.83 1.69 1.30 0.08
78 535 863 0.35 0.16 22.69 1.52 1.15 0.08
79 675 740 0.32 0.17 22.51 1.52 1.13 0.08
80 718 875 0.33 0.18 22.26 1.59 1.05 0.08
81 812 921 0.32 0.16 20.81 1.96 1.27 0.07
82 920 693 0.32 0.15 20.44 1.38 0.94 0.07
83 952 1049 0.33 0.12 18.44 2.67 1.60 0.06
84 1174 781 0.27 0.11 15.66 1.99 1.34 0.05
85 1130 1002 0.25 0.13 15.74 2.11 1.44 0.06
86 1146 1269 0.25 0.10 14.73 2.69 1.71 0.05
87 1213 1387 0.23 0.09 12.97 2.82 1.81 0.04
88 1228 1282 0.22 0.11 12.83 2.34 1.61 0.05
89 1298 1423 0.21 0.13 13.17 2.33 1.64 0.06
90 1306 1506 0.22 0.12 12.65 2.51 1.53 0.06
91 1352 1591 0.23 0.11 12.57 2.54 1.45 0.05
92 1423 1579 0.22 0.08 11.36 2.63 1.54 0.04
93 1607 1605 0.19 0.09 10.17 2.96 1.82 0.04
All years 1 8162 1167 0.27 0.13 16.87 2.18 1.43 0.06
5. Empirical results
Table 1 reports annual summary statistics for the total sample. The average
dividend payout ratio ranges from a high of 35% in 1976 and 1978 to a low of
19% in 1993. The average returnonequity ranges between 8 and 18%. The
average book value per share over the period is $16.87. The average P/B ratio is
2.18; however, this ratio is inﬂated by the presence of low book value ﬁrms in the
sample. Taking the inverse of the B/P ratio [shown as 1/Avg(B/P)] reduces the
average P/B ratio to 1.43. Finally, Table 1 reports the returnonasset ratio
[Avg ROA], which averages 6%. Collectively, these results illustrate the stability
of the key model parameters over our sample period.
292 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
Table 2
Annual crosssectional correlation of stock prices to book equity and EBO value measures
Table values represent crosssectional Spearman correlation coeﬃcients between the stock price on
June 30 of year t, book equity per share in calendar year t!1 (B), and two sets of three fundamental
value metrics computed using the Edwards—Bell—Ohlson (EBO) formula. The historical EBO value
measures use current year returnonequity (ROE) to proxy for future ROEs. The analyst based EBO
value measures use consensus I/B/E/S forecasts to proxy for future ROEs. Each set of nine value
measures diﬀers only in the assumed the number of forecasting periods (¹"1, 2, or 3). The discount
rate used is a threefactor industryspeciﬁc costofequity (Fama and French, 1997). All years
represents the timeseries mean of annual crosssectional correlations.
FF Threefactor FF Threefactor
¹"1 ¹"2 ¹"3 ¹"1 ¹"2 ¹"3
Year t Obs. B Historical EBO value
measures
Analyst based EBO value
measures
76 361 0.48 0.68 0.68 0.68 0.73 0.73 0.74
77 312 0.56 0.73 0.72 0.70 0.78 0.79 0.78
78 535 0.49 0.71 0.72 0.72 0.79 0.79 0.79
79 675 0.54 0.68 0.68 0.68 0.76 0.77 0.77
80 718 0.43 0.65 0.67 0.68 0.76 0.79 0.80
81 812 0.45 0.64 0.65 0.66 0.70 0.72 0.74
82 920 0.56 0.73 0.73 0.72 0.79 0.82 0.82
83 952 0.45 0.54 0.53 0.53 0.66 0.70 0.72
84 1174 0.69 0.71 0.69 0.68 0.81 0.82 0.82
85 1130 0.72 0.83 0.82 0.81 0.88 0.89 0.89
86 1146 0.69 0.76 0.74 0.74 0.87 0.87 0.87
87 1213 0.70 0.70 0.68 0.68 0.82 0.85 0.85
88 1228 0.74 0.77 0.75 0.74 0.87 0.89 0.88
89 1298 0.76 0.78 0.77 0.76 0.87 0.88 0.87
90 1306 0.67 0.74 0.73 0.72 0.84 0.86 0.87
91 1352 0.63 0.73 0.71 0.71 0.82 0.86 0.86
92 1423 0.64 0.64 0.61 0.60 0.81 0.83 0.83
93 1607 0.63 0.64 0.61 0.60 0.77 0.80 0.80
All years 18162 0.60 0.70 0.69 0.69 0.80 0.81 0.82
5.1. Correlation with stock prices
Table 2 reports crosssectional Spearman rank correlation coeﬃcients be
tween stock prices and either book value (B) or one of six EBO value metrics.
These six measures reﬂect the three diﬀerent empirical estimates of value
discussed earlier (Eqs. (3.1), (3.2) and (3.3)), estimated using historical earnings
and analyst forecasts. The discount rates used are industry speciﬁc costofequity
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 293
¹¹ We also used constant interest rates of 11%, 12%, and 13% as well as an industryspeciﬁc
singlefactor model. We ﬁnd varying the discount rate had little eﬀect on our results. Abarbanell and
Bernard (1995) also ﬁnd that allowing for intertemporal and ﬁrmspeciﬁc variations in r
C
had little
eﬀect on their results.
¹` Speciﬁcally, we use the risk premium for each industry reported in Fama and French (1997),
Table 7, plus a constant riskless rate of 0.0646 per year — the average annualized 30day tbill rate
over our sample period. The risk premiums we used are computed from 5year rolling regressions by
industry. See Fama and French (1997) for more details.
¹` We used a twoperiod model for »
because of concerns for the accuracy of the timeseries
earnings model beyond two years. However, using a threeperiod model for »
yields similar results.
based on a threefactor model (Fama and French, 1997).¹¹ The threefactor
model is estimated using mimicking portfolios for ﬁrm size and markettobook
ratios. In estimating these models, ﬁrms are grouped into 48 industry classes.
The threefactor industry costofequity ranges between 8.23% (for Alcoholic
Beverages) and 16.49% (for Real Estate).¹`
Over our sample period, book value (B) had an average correlation with price
of 0.60, suggesting that book equity explains around 36% of the crosssectional
variation in prices. Compared to B, each of the six value measures displays
a higher average correlation with stock price. »
explains around 49% of the
crosssectional variation in prices. Increasing ¹ from 1 to 3 produces slightly
weaker correlations, probably due to rapid decay in the precision of the ROE
forecasts over time. In sum, values based on historical ROEs contain important
valuerelevant information not captured by B.
Table 2 also shows that the crosssectional correlation with price increases
when EBO value is estimated using analyst forecasts. Over our sample period,
»
explains around twothirds of the crosssectional variation in prices. Increas
ing ¹ from 1 to 3 produces slightly better correlations, but varying the discount
rate again has little eﬀect. Evidently, analysts’ earnings forecasts contain more
valuerelevant information than is reﬂected in historical ROEs. The superiority
of analysts over simple randomwalk models in forecasting earnings is well
documented (Fried and Givoly, 1982; Brown et al., 1987a,b; O’Brien, 1988). Our
ﬁndings suggest analyst forecasts also better reﬂect the market expectations of
earnings implicit in the EBO model.
Table 2 suggests that using all variations of the estimated » metric is unnec
essary. Therefore, for the remainder of this paper, we use »
to denote the
fundamental value computed from historical ROEs, a threefactor industry
speciﬁc discount rate, and a forecast horizon of twoperiod (Eq. (3.2)). We use
»
to denote the fundamental value computed using the mean analyst forecast,
a three factor industryspeciﬁc discount rate, and a forecast horizon of three
periods (Eq. (3.3)).
13
294 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
5.2. Correlation with future returns: unidimensional analyses
The main focus of this study is on the prediction of future returns. As a ﬁrst
step, we construct unidimensional portfolios based on market value of
equity (ME), B/P, and »
/P. Table 3 reports characteristics of quintile portfolios
formed on the basis of these ﬁrm characteristics. This table is constructed
by sorting all sample ﬁrms into quintiles at the end of June each year.
Firm size quintiles are formed in two ways. First, as in Fama and French (1992),
the size decile cutoﬀs are based on June 30 prices for all NYSE ﬁrms. Second,
we use annual insample ﬁrm size cutoﬀs to form quintiles. For each
portfolio, Table 3 reports the average B/P, ME, and »
/P values, as well
as the average postranking market betas, and average buyandhold return
over the next 12 months (Ret12), 24 months (Ret24), and 36 months
(Ret36). Market beta for each ﬁrm is estimated using an equalweighted
market index and each ﬁrm’s monthly returns over the next 36 months.
The last row in each panel shows the number of ﬁrmyear observations
in each portfolio, and applies to all variables except the stock return
variables. When we require availability of stock returns, the number of
observations drop to 16,549, 14,385, and 12,377 for Ret12, Ret24 and Ret36,
respectively.
The right column of Table 3 reports the diﬀerences in means between the top
(Q5) and bottom (Q1) quintiles. The statistical signiﬁcance of this diﬀerence is
assessed using a Monte Carlo simulation technique similar to those discussed in
Barber and Lyon (1997), Kothari and Warner (1997), and Lyon et al. (1998).
Speciﬁcally, we form empirical reference distributions by randomly assigning
the population of eligible ﬁrms each year into quintile portfolios (without
replacement). This procedure generates ﬁve random quintile portfolios each
year with the same number of observations as the actual quintile portfolios. We
repeat the process until we have obtained 1000 sets of quintile portfolios for each
year. We then compute the mean returns for the Q5—Q1 portfolio. To determine
statistical signiﬁcance, we use pvalues calculated from the simulated empirical
distribution of mean Q5—Q1 returns.
Our randomization procedure avoids the three main econometric problems
discussed in Lyon et al. (1998) and Kothari and Warner (1997). First,
our reference portfolios only contain ﬁrms that are available for investing
at the same time as our sample ﬁrms. This avoids the new listing or
survivor bias. Second, we compute returns for the reference portfolios in
exactly the same manner as for the actual portfolios (that is, both reﬂect
buyandhold returns over the same time horizon). This avoids the re
balancing bias, and adjusts for serial correlations in returns induced by
overlapping holding periods. Finally, the use of pvalues calculated from
the simulated empirical distribution avoids the skewness bias discussed in the
literature.
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 295
Table 3
Characteristics of quintileportfolios formed by ME, B/P, and »
/P
This table reports the characteristics of quintile portfolios formed at the end of June each year by
market value of equity (ME), bookvaluetoprice (B/P), and analyst based EBO valuetoprice
(»
/P). Each panel reports mean values for individual quintile characteristics. ME is the market
value of shareholders’ equity as of June 30 of year t, expressed in millions. ME quintiles are based on
size cutoﬀs for all NYSE ﬁrms (Panel A) and on insample size cutoﬀs (Panel B). Book value (B) is
book equity per share in calendar year t!1. Price (P) is the stock price at the end of June in year t.
Analyst based EBO value (»
) is a fundamental value measure derived using current I/B/E/S
consensus analyst predictions of future earnings available prior to June 30 of year t. beta is estimated
using monthly returns over the 36 months beginning July of year t. Ret12, Ret24, and Ret36 are the
average oneyear, twoyear, threeyear buyandhold return for the portfolio. Obs. is the number of
observations in each quintile and applies to all variables except Ret12, Ret24, and Ret36. Results in
the All Firms column represent unconditional means. Q5!Q1 Diﬀ. results represent diﬀerences in
means between the top (Q5) and bottom(Q1) quintiles. The statistical signiﬁcance of this diﬀerence is
derived using Monte Carlo simulation. Speciﬁcally, we form empirical reference distributions by
randomly assigning eligible ﬁrms into quintiles each year. ***, **, * signify that the observed
diﬀerence between the extreme quintiles is signiﬁcantly diﬀerent from those of the reference
distribution at the 1%, 5% and 10%levels, respectively (twotailed). The sample period is 1977—1992
(t"77 to 92).
Panel A — Marketequity portfolios (NYSE size quintiles)
Q1
(Low ME)
Q2 Q3 Q4 Q5
(High ME)
All
Firms
Q5!Q1
Diﬀ.
ME 9 25 59 157 2293 1230
B/P 1.39 1.04 0.79 0.66 0.62 0.69 !0.77
»
/P 1.33 1.04 0.92 0.87 0.91 0.91 !0.42
Beta 0.73 1.06 1.14 1.17 1.03 1.08 0.30
Ret12 0.379 0.239 0.153 0.159 0.146 0.159 !0.233***
Ret24 0.515 0.340 0.304 0.312 0.305 0.311 !0.210*
Ret36 0.835 0.556 0.442 0.487 0.497 0.493 !0.338*
Obs. 232 1144 2692 4761 9333 1 8162
Panel B — Marketequity portfolios (insample size quintiles)
Q1
(Low ME)
Q2 Q3 Q4 Q5
(High ME)
All
Firms
Q5!Q1
Diﬀ.
ME 41 117 277 722 4983 1230
B/P 0.92 0.69 0.61 0.65 0.60 0.69 !0.32
»
/P 1.00 0.90 0.85 0.87 0.95 0.91 !0.05
Beta 1.08 1.17 1.11 1.07 0.98 1.08 !0.10
Ret12 0.158 0.157 0.161 0.158 0.159 0.159 0.001
Ret24 0.285 0.311 0.319 0.310 0.330 0.311 0.045**
Ret36 0.459 0.486 0.489 0.503 0.525 0.493 0.066**
Obs 3622 3636 3632 3632 3640 1 8162
296 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
Table 3 (continued)
Panel C — Booktoprice (B/P) Portfolios
Q1
(Low B/P)
Q2 Q3 Q4 Q5
(High B/P)
All
Firms
Q5!Q1
Diﬀ.
B/P 0.24 0.42 0.60 0.81 1.39 0.69 —
ME 1641 1434 1345 1068 666 1230 !975
»
/P 0.75 0.86 0.93 1.02 1.01 0.91 0.260
Beta 1.29 1.17 1.05 0.93 0.97 1.08 !0.320
Ret12 0.137 0.148 0.156 0.166 0.186 0.159 0.049***
Ret24 0.251 0.300 0.332 0.338 0.333 0.311 0.082***
Ret36 0.407 0.450 0.513 0.535 0.558 0.493 0.151***
Obs 3621 3628 3634 3636 3643 1 8162
Panel D — »
/P portfolios
Q1
(Low »
/P)
Q2 Q3 Q4 Q5
(High »
/P)
All
Firms
Q5!Q1
Diﬀ.
»
/P 0.40 0.70 0.87 1.06 1.54 0.91 —
B/P 0.60 0.59 0.68 0.75 0.85 0.69 0.25
ME 812 1252 1531 1377 1177 1230 365
Beta 1.24 1.09 1.05 0.99 1.03 1.08 !0.210
Ret12 0.138 0.154 0.159 0.172 0.169 0.159 0.031***
Ret24 0.217 0.298 0.317 0.351 0.369 0.311 0.152***
Ret36 0.331 0.450 0.491 0.549 0.637 0.493 0.306***
Obs 3626 3632 3632 3632 3640 1 8162
¹" We thank the referee for his insights on this point.
Despite these advantages, this simulation procedure still has a potential
deﬁciency.¹" By randomly assigning ﬁrms to quintiles 1 through 5, the proced
ure creates portfolios whose covariance is equal to the average covariance across
all returns in our sample. If the actual correlation structure between portfolios
one and ﬁve diﬀers signiﬁcantly from this average covariance, the pvalues based
on the simulation may be misleading. In particular, crosscorrelations in the
data may cause the variance of the simulated Q5—Q1 returns to be lower than
that of the true Q5—Q1 returns. To mitigate this problem, we will provide
additional corroborating evidence using more traditional statistical procedures
(see Tables 8 and 9).
Panels A and B examine the ME eﬀect for our sample. Panel A shows that
a smallﬁrm eﬀect exists when NYSE size quintiles are used. Over 12, 24, and
36 month periods following portfolio formation, small ﬁrms generally outper
form large ﬁrms. However, because we require that ﬁrms be followed by
analysts, larger ﬁrms dominate the sample — the last row of Panel A shows that
over 80% of our ﬁrms are larger than the median NYSE ﬁrm. Panel B shows
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 297
¹` Kothari et al. (1995) show annual betas are more highly correlated with crosssectional returns.
¹" Results for »
/P are similar to those for »
/P. Interestingly, returns from the B/P strategy
exhibit a seasonal pattern not found in the other two strategies — Fig. 1 shows that the January eﬀect
is much more pronounced in the B/P strategy than in the »/P strategy.
that when our ﬁrms are grouped by insample size cutoﬀs, large ﬁrms actually
outperform small ﬁrms over 24 and 36 month holding periods. Thus, in our
sample, the ﬁrmsize eﬀect is driven primarily by a small set of the smallest
stocks.
Panel C conﬁrms a B/P eﬀect for our sample. The lowest B/P quintile ﬁrms
earn an average return of 13.7% over the next 12months while the highest B/P
quintile ﬁrms earn 18.6%. The diﬀerence of 4.9% is statistically signiﬁcant at
1%, and is comparable in magnitude to other studies reporting the B/P eﬀect
using I/B/E/Sconstrained samples (e.g., Dechow and Sloan, 1997). The B/P
eﬀect is also seen over longer holding periods. The relation between B/P and
future returns is generally monotonic across the quintiles. As in Lakonishok et
al. (1994), we ﬁnd low B/P ﬁrms have higher betas than high B/P ﬁrms. This
result suggests that the B/P eﬀect is not due to diﬀerences in market risk. High
B/P betas may also be biased downward due to nonsychroneity, although the
larger nature of our sample ﬁrms may reduce this problem.¹`
Panel D shows that »
/P portfolios have some similarities with B/P port
folios. »
/P and B/P are positively correlated. High »
/P ﬁrms, like high B/P
ﬁrms, tend to have lower market betas. Moreover, these results show that »
/P
also predicts returns. The shortterm prediction results for »
/P are slightly
weaker than the results for B/P. The lowest »
/P quintile ﬁrms earn 13.8% over
the next 12 months, while the highest »
/P quintile ﬁrms earn only 16.9%.
Furthermore, the returns pattern is not monotonic across the quintiles. How
ever, over 24 and 36 months, high »
/P ﬁrms signiﬁcantly outperform low »
/P
ﬁrms. Indeed, over these longer horizon, we observe a monotonic relation in
returns across the quintiles. Over 36 months, for example, the spread between
the highest and lowest »
/P portfolios is 30.6%.
Fig. 1 illustrates the cumulative returns from a 36 month buyandhold
strategy involving B/P and »
/P. To construct this graph, a longposition is
taken in the top quintile ﬁrms based on each ratio, and a shortposition is taken
in the bottom quintile ﬁrms. This graph reports the diﬀerence in cumulative
buyandhold returns between the top and bottom quintiles at monthly intervals
over the next three years. The graph shows that over a 36 month period, the »/P
strategy outperforms the B/P strategy by a wide margin.¹"
In summary, we ﬁnd that »
/P is much better at explaining crosssectional
prices than B/P. »
/P is also a better predictor of longterm returns. However,
»
/P is not necessarily more useful for predicting returns over 12 month
intervals. Our ﬁndings suggest that while analyst consensus forecasts provide
298 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
Fig. 1. Cumulative (buyandhold) returns produced by B/P and »
/P Trading strategies. This
ﬁgure shows the cumulative (buyandhold) returns from B/P and »
/P based trading strategies. B is
book equity per share in calendar year t!1. P is price per share at the end of June 30 of year t. »
is
a fundamental value estimate based on the consensus analyst forecast as of May of year t. Each year,
portfolios are formed at the end of June by sorting ﬁrms into quintiles on the basis of B/P and »
/P.
For each investment strategy, this graph depicts the cumulative buyandhold returns produced by
buying ﬁrms in the top quintile and selling ﬁrms in the bottom quintile at the beginning of July, and
maintaining these investments until the end of the indicated month. The sample period is 1979—1991
(year t"1979 to 1991).
a good proxy for market expectations, trading on the basis of these forecasts
does not necessarily yield higher shortrun (12 month) returns than trading on
B/P. In later tests, we explore a strategy that seeks to improve on the consensus
earnings forecast.
5.3. Correlation with future returns: bidimensional analyses
We now consider how much of the explanatory power of »
/P for longterm
returns is due to its correlation with ﬁrm size and B/P. Fama and French (1992)
show that both ﬁrm size and B/P have predictive power for crosssectional
returns. We examine the extent to which these two factors explain the predictive
power of »
/P.
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 299
T
a
b
l
e
4
A
v
e
r
a
g
e
3
6

m
o
n
t
h
b
u
y

a
n
d

h
o
l
d
r
e
t
u
r
n
s
f
o
r
b
i

d
i
m
e
n
s
i
o
n
a
l
p
o
r
t
f
o
l
i
o
s
T
h
i
s
t
a
b
l
e
r
e
p
o
r
t
s
t
h
e
a
v
e
r
a
g
e
3
6

m
o
n
t
h
r
e
a
l
i
z
e
d
r
e
t
u
r
n
s
f
o
r
b
i

d
i
m
e
n
s
i
o
n
a
l
p
o
r
t
f
o
l
i
o
s
,
f
o
r
m
e
d
u
s
i
n
g
t
w
o
e
x
a
n
t
e
ﬁ
r
m
c
h
a
r
a
c
t
e
r
i
s
t
i
c
s
s
i
m
u
l
t
a
n
e
o
u
s
l
y
.
O
n
J
u
n
e
3
0
o
f
y
e
a
r
t
,
a
l
l
s
a
m
p
l
e
ﬁ
r
m
s
a
r
e
s
o
r
t
e
d
(
i
n
d
e
p
e
n
d
e
n
t
l
y
)
i
n
t
o
q
u
i
n
t
i
l
e
s
a
c
c
o
r
d
i
n
g
t
o
t
h
e
v
a
r
i
a
b
l
e
s
d
e
s
i
g
n
a
t
e
d
i
n
t
h
e
p
a
n
e
l
s
.
B
o
o
k
v
a
l
u
e
(
B
)
i
s
b
o
o
k
e
q
u
i
t
y
p
e
r
s
h
a
r
e
i
n
c
a
l
e
n
d
a
r
y
e
a
r
t
!
1
.
P
r
i
c
e
(
P
)
i
s
t
h
e
s
t
o
c
k
p
r
i
c
e
a
t
t
h
e
e
n
d
o
f
J
u
n
e
i
n
y
e
a
r
t
.
A
n
a
l
y
s
t

b
a
s
e
d
E
B
O
v
a
l
u
e
(
»
)
i
s
a
f
u
n
d
a
m
e
n
t
a
l
v
a
l
u
e
m
e
a
s
u
r
e
d
e
r
i
v
e
d
u
s
i
n
g
I
/
B
/
E
/
S
c
o
n
s
e
n
s
u
s
a
n
a
l
y
s
t
f
o
r
e
c
a
s
t
s
t
o
p
r
o
x
y
f
o
r
f
u
t
u
r
e
e
a
r
n
i
n
g
s
.
T
a
b
l
e
v
a
l
u
e
s
r
e
p
r
e
s
e
n
t
m
e
a
n
b
u
y

a
n
d

h
o
l
d
r
e
t
u
r
n
s
f
o
r
e
a
c
h
p
o
r
t
f
o
l
i
o
o
v
e
r
t
h
e
n
e
x
t
3
6
m
o
n
t
h
s
.
T
h
e
n
u
m
b
e
r
o
f
o
b
s
e
r
v
a
t
i
o
n
s
i
s
s
h
o
w
n
i
n
p
a
r
e
n
t
h
e
s
e
s
.
R
e
s
u
l
t
s
f
o
r
A
l
l
F
i
r
m
s
r
e
p
r
e
s
e
n
t
u
n
c
o
n
d
i
t
i
o
n
a
l
m
e
a
n
s
.
T
h
e
Q
5
!
Q
1
d
i
ﬀ
.
r
e
s
u
l
t
s
r
e
p
r
e
s
e
n
t
d
i
ﬀ
e
r
e
n
c
e
s
i
n
m
e
a
n
r
e
t
u
r
n
s
b
e
t
w
e
e
n
t
h
e
Q
5
a
n
d
Q
1
q
u
i
n
t
i
l
e
s
,
c
o
n
t
r
o
l
l
i
n
g
f
o
r
q
u
i
n
t
i
l
e
m
e
m
b
e
r
s
h
i
p
i
n
o
t
h
e
r
v
a
r
i
a
b
l
e
.
T
h
e
s
t
a
t
i
s
t
i
c
a
l
s
i
g
n
i
ﬁ
c
a
n
c
e
o
f
t
h
i
s
d
i
ﬀ
e
r
e
n
c
e
i
s
d
e
r
i
v
e
d
u
s
i
n
g
a
M
o
n
t
e
C
a
r
l
o
s
i
m
u
l
a
t
i
o
n
t
e
c
h
n
i
q
u
e
.
S
p
e
c
i
ﬁ
c
a
l
l
y
,
w
e
f
o
r
m
e
m
p
i
r
i
c
a
l
r
e
f
e
r
e
n
c
e
d
i
s
t
r
i
b
u
t
i
o
n
s
b
y
r
a
n
d
o
m
l
y
a
s
s
i
g
n
i
n
g
e
l
i
g
i
b
l
e
ﬁ
r
m
s
i
n
t
o
q
u
i
n
t
i
l
e
s
e
a
c
h
y
e
a
r
w
h
i
l
e
h
o
l
d
i
n
g
q
u
i
n
t
i
l
e
m
e
m
b
e
r
s
h
i
p
i
n
t
h
e
o
t
h
e
r
v
a
r
i
a
b
l
e
c
o
n
s
t
a
n
t
.
*
*
*
,
*
*
,
*
s
i
g
n
i
f
y
t
h
a
t
t
h
e
o
b
s
e
r
v
e
d
d
i
ﬀ
e
r
e
n
c
e
b
e
t
w
e
e
n
t
h
e
Q
5
a
n
d
Q
1
q
u
i
n
t
i
l
e
s
i
s
s
i
g
n
i
ﬁ
c
a
n
t
l
y
d
i
ﬀ
e
r
e
n
t
f
r
o
m
t
h
o
s
e
o
f
t
h
e
r
e
f
e
r
e
n
c
e
d
i
s
t
r
i
b
u
t
i
o
n
a
t
t
h
e
1
%
,
5
%
a
n
d
1
0
%
l
e
v
e
l
s
,
r
e
s
p
e
c
t
i
v
e
l
y
(
o
n
e

t
a
i
l
e
d
)
.
T
h
e
s
a
m
p
l
e
p
e
r
i
o
d
i
s
1
9
7
7
—
1
9
9
1
(
t
"
7
7
t
o
9
1
)
.
P
a
n
e
l
A
—
A
v
e
r
a
g
e
3
6

m
o
n
t
h
b
u
y

a
n
d

h
o
l
d
r
e
t
u
r
n
s
f
o
r
p
o
r
t
f
o
l
i
o
s
f
o
r
m
e
d
o
n
t
h
e
b
a
s
i
s
o
f
b
o
t
h
ﬁ
r
m
s
i
z
e
(
i
n

s
a
m
p
l
e
q
u
i
n
t
i
l
e
s
)
a
n
d
»
/
P
A
n
a
l
y
s
t

b
a
s
e
d
E
B
O
v
a
l
u
e

t
o

m
a
r
k
e
t
(
»
/
P
)
Q
1
(
L
o
w
»
/
P
)
Q
2
Q
3
Q
4
Q
5
(
H
i
g
h
»
/
P
)
A
l
l
F
i
r
m
s
Q
5
!
Q
1
D
i
ﬀ
.
S
i
z
e
q
u
i
n
t
i
l
e
s
Q
1
(
S
m
a
l
l
M
E
)
0
.
3
1
9
0
.
3
8
3
0
.
4
1
8
0
.
5
2
5
0
.
5
9
0
0
.
4
5
9
0
.
2
7
1
*
*
*
4
9
7
(
3
7
8
)
(
4
0
5
)
(
4
2
5
)
(
6
6
8
)
(
2
3
7
3
)
Q
2
0
.
2
8
7
0
.
5
2
0
0
.
5
3
9
0
.
5
0
4
0
.
5
7
7
0
.
4
8
6
0
.
2
9
0
*
*
*
(
4
8
0
)
(
5
0
0
)
(
5
1
9
)
(
4
4
7
)
(
4
6
0
)
(
2
4
0
6
)
Q
3
0
.
3
8
3
0
.
4
2
7
0
.
4
8
6
0
.
5
5
7
0
.
6
5
3
0
.
4
8
9
0
.
2
7
0
*
*
*
(
5
5
9
)
(
5
1
8
)
(
4
4
7
)
(
3
6
0
)
(
2
3
9
2
)
(
5
0
8
)
Q
4
0
.
3
1
8
0
.
4
7
3
0
.
4
9
8
0
.
5
6
5
0
.
7
0
6
0
.
5
0
3
0
.
3
8
8
*
*
*
(
5
2
9
)
(
5
3
5
)
(
5
1
1
)
(
5
5
4
)
(
4
0
7
)
(
2
5
3
6
)
Q
5
(
L
a
r
g
e
M
E
)
0
.
3
5
0
0
.
4
3
3
0
.
4
9
6
0
.
5
7
8
0
.
6
7
9
0
.
5
2
5
0
.
3
2
9
*
*
*
(
3
7
0
)
(
4
9
8
)
(
5
4
2
)
(
6
6
5
)
(
5
9
5
)
(
2
6
7
0
)
A
l
l
ﬁ
r
m
s
0
.
3
3
1
0
.
4
5
0
0
.
4
9
1
0
.
5
4
9
0
.
6
3
7
0
.
4
9
3
(
2
3
8
4
)
(
2
4
7
0
)
(
2
4
9
5
)
(
2
5
3
8
)
(
2
4
9
0
)
(
1
2
3
7
7
)
Q
5
!
Q
1
d
i
ﬀ
.
0
.
0
3
1
0
.
0
5
0
0
.
0
7
8
*
0
.
0
5
3
*
0
.
0
8
9
*
300 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
P
a
n
e
l
B
—
A
v
e
r
a
g
e
3
6

m
o
n
t
h
b
u
y

a
n
d

h
o
l
d
r
e
t
u
r
n
s
f
o
r
p
o
r
t
f
o
l
i
o
s
f
o
r
m
e
d
o
n
t
h
e
b
a
s
i
s
o
f
b
o
t
h
B
/
P
a
n
d
»
/
P
A
n
a
l
y
s
t

b
a
s
e
d
E
B
O
v
a
l
u
e

t
o

m
a
r
k
e
t
(
»
/
P
)
Q
1
(
L
o
w
»
/
P
)
Q
2
Q
3
Q
4
Q
5
(
H
i
g
h
»
/
P
)
A
l
l
F
i
r
m
s
Q
5
!
Q
1
D
i
ﬀ
.
B
o
o
k

t
o

m
a
r
k
e
t
Q
1
(
L
o
w
B
/
P
)
0
.
3
1
6
0
.
4
6
8
0
.
3
4
2
0
.
4
5
7
0
.
6
3
4
0
.
4
0
7
0
.
3
1
8
*
*
*
(
9
9
8
)
(
5
9
2
)
(
2
9
6
)
(
2
6
4
)
(
2
6
9
)
(
2
4
1
9
)
Q
2
0
.
3
6
6
0
.
4
6
1
0
.
4
8
9
0
.
4
1
5
0
.
5
1
6
0
.
4
5
0
0
.
1
5
0
*
*
*
(
4
9
5
)
(
6
9
4
)
(
5
7
3
)
(
3
3
3
)
(
3
4
4
)
(
2
4
3
9
)
Q
3
0
.
3
9
6
0
.
4
4
0
0
.
5
3
0
0
.
5
7
6
0
.
5
6
6
0
.
5
1
3
0
.
1
7
0
*
(
2
9
5
)
(
5
1
5
)
(
6
6
0
)
(
5
6
5
)
(
4
5
3
)
(
2
4
8
8
)
Q
4
0
.
3
5
0
0
.
4
2
2
0
.
4
8
4
0
.
5
8
9
0
.
6
3
0
0
.
5
3
5
0
.
2
8
0
*
*
*
(
2
1
0
)
(
3
4
1
)
(
5
3
3
)
(
8
6
6
)
(
6
0
0
)
(
2
5
5
0
)
Q
5
(
h
i
g
h
B
/
P
)
0
.
2
6
3
0
.
4
4
2
0
.
5
4
4
0
.
5
8
8
0
.
7
3
2
0
.
5
5
8
0
.
4
6
9
*
*
*
(
3
8
6
)
(
3
2
8
)
(
4
3
3
)
(
5
1
0
)
(
8
2
4
)
(
2
4
8
1
)
A
l
l
F
i
r
m
s
0
.
3
3
1
0
.
4
5
0
0
.
4
9
1
0
.
5
4
9
0
.
6
3
7
0
.
4
9
3
(
2
3
8
4
)
(
2
4
7
0
)
(
2
4
9
5
)
(
2
5
3
8
)
(
2
4
9
0
)
(
1
2
3
7
7
)
Q
5
!
Q
1
D
i
ﬀ
.
!
0
.
0
5
3
!
0
.
0
2
6
0
.
2
0
2
*
*
*
0
.
1
3
1
0
.
0
9
8
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 301
¹` Results are similar when we use NYSEsize quintiles, however some cells have very few
observations.
¹` When a ﬁrm drops out over our holding period, a terminating return to the delisting date is
computed. The proceeds from termination, if any, are equally assigned to surviving ﬁrms in the same
portfolio. Firms that switch exchanges are traced to their new exchange listings and retained in their
original portfolios.
To address this question, we examine future returns to »
/P portfolios while
controlling for ME and B/P. Table 4 reports the average realized return to
36month buyandhold strategies for bidimensional portfolios. To construct
this table, we independently sort ﬁrms into quintiles based on each partitioning
variable as of the end of June each year. Stocks are then assigned to one of 25
portfolios based on their bidimensional ranking. Panel A reports portfolio
returns for »
/P and ﬁrm size (using insample size cutoﬀs).¹` Panel B reports
portfolio returns for »
/P and B/P.
We again assess the statistical signiﬁcance of the diﬀerence between Q1 and
Q5 portfolios using a Monte Carlo simulation technique. In this analysis, we
hold quintile membership in the other variable constant while randomizing
across the variable of interest. For example, to create the empirical distribution
for »
/P in Panel B, we assigned the total population of ﬁrms within each B/P
quintile into random »
/P quintiles each year (without replacement). This
procedure controls for B/P membership each year as well as any serial correla
tion in yeartoyear returns. The resulting empirical distribution allows us to
assess the incremental usefulness of one variable in predicting returns after
controlling for the other.¹`
Panel A shows that »
/P has strong predictive power in all ﬁve size quintiles.
The diﬀerence in Q5—Q1 returns range from 27.0% to 38.8%. The right column
shows that these diﬀerences are statistically signiﬁcant at the 1% level in each of
ﬁve size quintiles. The bottom row shows that when ﬁrms are divided using
insample size cutoﬀs, large ﬁrms slightly outperform small ﬁrms after control
ling for »
/P.
Panel B shows the interaction of the B/P and »
/P eﬀects. Looking down each
column, we see that the B/P eﬀect is much weaker, and no longer monotonic
after controlling for »
/P. Conversely, looking across each row, we observe
a largely monotonic and statistically signiﬁcant relation between »
/P and
returns. The simulation results show »
/P explains longrun returns within all
ﬁve B/P portfolios; however, the B/P eﬀect survives in only one »
/P quintile.
Taken together, Panels A and B suggest that in longer time horizons, the
predictive power of »
/P for future returns is not explained by either B/P or ﬁrm
size. Later, we examine the relative contribution of each variable in returns
prediction using a multiple regression approach.
302 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
¹" Similar arguments have been made in Dechow and Sloan (1995), Daniel and Mande (1994),
LaPorta (1996), LaPorta et al. (1997). Unique features of our study include the use of a comprehens
ive valuation model that utilizes earnings forecasts the development of a prediction model for
analyst forecast errors, as well as a trading strategy that directly exploits the predicted error in
analysts.
`" ¸tg was not reported by I/B/E/S until 1981. In the pre1981 period, we estimated the longterm
growth rate using the growth rate implicit between F½1 and F½2.
5.4. The relation between analyst forecast errors and ex ante
ﬁrm characteristics
In this part of the paper, we investigate the quality of I/B/E/S consensus
forecasts. This investigation has two motivations. First, the reliability of
»
depends critically on the quality of the earnings forecast used. In using
I/B/E/S consensus forecasts to estimate »
, we implicitly assume that these
forecasts are unbiased with respect to public information. If this assumption is
not true, we should be able to further improve the ability of »
/P to predict
returns by incorporating the predictable errors. Second, the mispricing hypothe
sis suggests a relation between certain characteristics and the direction of
subsequent forecast errors, while the risk hypothesis does not. Therefore, this
investigation should be helpful in distinguishing between the two hypotheses.¹"
Speciﬁcally, we investigate the relation between analyst forecast errors and
four ex ante ﬁrm characteristics — the booktomarket ratio (B/P), past sales
growth (SG), analyst consensus longterm earnings growth forecast (¸tg), and
a new measure we call OP (for analyst optimism). The use of SG is suggested by
Lakonishok et al. (1994)’s [LSV] ﬁnding that ﬁrms with higher (lower) past sales
growth earn lower (higher) subsequent returns. Like LSV, we deﬁne SG in terms
of the percentage growth in sales over the past ﬁve years. LSV argue that their
ﬁnding is due to investor over optimism(pessimism) in ﬁrms with high (low) past
sales growth. Thus, the mispricing hypothesis predicts that high (low) SGs are
associated with over optimistic (pessimistic) I/B/E/S forecasts.
The use of the consensus longterm earnings growth forecast (¸tg) is moti
vated by LaPorta (1996) and Dechow and Sloan (1997). LaPorta shows that
ﬁrms with higher longterm earnings forecasts (high ¸tg ﬁrms) tend to earn
lower subsequent returns. Dechow and Sloan (1997) show that the Ltg eﬀect
accounts for a signiﬁcant portion of the return to contrarian investment stra
tegies, including strategies based on the B/P and E/P ratios. We extend this
literature by examining the power of Ltg to predict errors in longterm analyst
forecasts, alone and in combination with variables.`"
OP is a measure of analyst optimism derived from EBO fundamental value
measures. Speciﬁcally, OP"(»
!»
)/"»
". OP measures the extent to which
equity values based on analyst forecasts deviate from similar valuations based
on historical earnings. Past studies show that analysts are more accurate than
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 303
Table 5
Accounting proﬁtability of quintile portfolios formed by B/P, SG, ¸tg, and OP
This table reports the accounting proﬁtability characteristics of quintile portfolios formed at the end
of June each year. B/P is the booktomarket ratio, where B is the book value per share for the ﬁscal
year ended in year t!1, and P is the stock price at the end of June in year t. SG is the ﬁveyear
growth rate in sales from period t!6 to t!1. ¸tg is the consensus longterm earnings growth
forecast from I/B/E/S as of May in year t. OP"(»
!»
)/"»
", where »
is an EBO value derived
using current I/B/E/S consensus forecasts and »
is a similar EBO value measure derived using
historical ROEs. OP measures the extent to which equity values based on analyst forecasts deviate
from similar valuations based on historical earnings. Each panel reports mean values for individual
quintiles. Ret36 is the average threeyear buyandhold return for the portfolio. ROE
G
is equal to
reported net income in year i divided by the average of year i and i!1 book values. The forecast
error for each observation (FErr
R>G
) is computed by subtracting the forecasted (FROE
R>G
) from the
actual reported ROE in period t#i. To be included, ﬁrms are required to have all of the above
variables available. Obs. is the number of observations. Results for All Firms represent uncondi
tional means. Q5!Q1 Diﬀ. results represent diﬀerences in means between the top (Q5) and bottom
(Q1) quintiles. The statistical signiﬁcance of this diﬀerence is derived using Monte Carlo simulation.
Speciﬁcally, we form empirical reference distributions by randomly assigning eligible ﬁrms into
quintiles each year. ***, **, * signify that the observed diﬀerence between the extreme quintiles is
signiﬁcantly diﬀerent from those of the reference distribution at the 1%, 5% and 10% levels,
respectively (onetailed). Sample period is 19771991 (t"77 to 91).
Panel A — Booktomarket (B/P) portfolios
Q1
(Low B/P)
Q2 Q3 Q4 Q5
(High B/P)
All
Firms
Q5!Q1
Diﬀ.
B/P 0.265 0.455 0.642 0.846 1.366 0.716 —
ROE
R¹
0.202 0.157 0.124 0.108 0.053 0.129 !0.149
FROE
R
0.285 0.200 0.158 0.131 0.080 0.170 !0.205
FROE
R>¹
0.289 0.208 0.171 0.144 0.104 0.183 !0.185
FROE
R>`
0.280 0.207 0.173 0.147 0.107 0.183 !0.173
FErr
R
0.037 0.025 0.028 0.031 0.060 0.036 0.023***
FErr
R>¹
0.054 0.058 0.062 0.051 0.073 0.059 0.019**
FErr
R>`
0.113 0.051 0.066 0.049 0.074 0.070 !0.039**
Ret36 0.462 0.530 0.543 0.617 0.613 0.553 0.151***
Obs. 2350 2378 2372 2375 2386 11861
earnings forecasts based on timeseries models. We ﬁnd the same is true for our
sample of ﬁrms. However, in the crosssection, analyst forecasts that deviate the
most from historical earnings may reﬂect an underweighting of historical informa
tion. OP captures these deviations. Speciﬁcally, OP captures analyst optimism
relative to past reported ROEs — analysts expect the highest (lowest) OP ﬁrms to
experience the most (least) ROE growth. If analysts underweight historical proﬁtab
ility benchmarks in making their forecasts, then OP will be positively correlated
with analyst optimism. Two examples from the behavioral literature that discuss
this possibility are Tversky and Kahneman (1984) and DeBondt (1993).
Table 5 reports characteristics of quintile portfolios formed at the end of June
each year by B/P, SG, OP, and ¸tg. ROE
R¹
is the actual reported returnsonequity
304 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
Panel B — Sales growth (SG) portfolios
Q1
(Low SG)
Q2 Q3 Q4 Q5
(High SG)
All
Firms
Q5!Q1
Diﬀ.
SG !0.014 0.451 0.787 1.322 9.081 2.331 —
ROE
R¹
0.087 0.125 0.143 0.148 0.141 0.129 0.054
FROE
R
0.132 0.161 0.176 0.185 0.199 0.170 0.067
FROE
R>¹
0.154 0.174 0.186 0.192 0.207 0.183 0.053
FROE
R>`
0.155 0.174 0.185 0.192 0.207 0.183 0.052
FErrt 0.034 0.024 0.027 0.031 0.067 0.036 0.033***
FErr
R>¹
0.060 0.051 0.046 0.051 0.092 0.059 0.032***
FErr
R>`
0.066 0.057 0.060 0.053 0.118 0.070 0.052**
Ret36 0.637 0.626 0.558 0.556 0.384 0.553 0.253***
Obs. 2365 2372 2372 2372 2380 11861
Panel C — Analyst optimism (OP) portfolios
Q1
(Low OP)
Q2 Q3 Q4 Q5
(High OP)
All
Firms
Q5!Q1
Diﬀ.
OP !0.018 0.125 0.283 0.646 5.23 1.225 —
ROE
R¹
0.179 0.166 0.162 0.120 0.017 0.129 !0.162
FROE
R
0.152 0.179 0.189 0.176 0.156 0.170 0.002
FROE
R>¹
0.148 0.178 0.194 0.192 0.201 0.183 0.053
FROE
R>`
0.147 0.177 0.194 0.193 0.202 0.183 0.055
FErr
R
0.025 0.016 0.022 0.038 0.081 0.036 0.056***
FErr
R>¹
0.050 0.030 0.041 0.052 0.126 0.059 0.076***
FErr
R>`
0.062 0.059 0.046 0.074 0.114 0.070 0.052***
Ret36 0.554 0.643 0.591 0.521 0.455 0.553 !0.099***
Obs. 2365 2372 2372 2372 2380 11861
Panel D — Longterm growth forecast (¸tg) portfolios
Q1
(Low ¸tg)
Q2 Q3 Q4 Q5
(High ¸tg)
All
Firms
Q5!Q1
Diﬀ.
¸tg 0.046 0.105 0.135 0.169 0.306 0.153 —
ROE
R¹
0.105 0.122 0.145 0.147 0.126 0.129 0.021
FROE
R
0.135 0.155 0.179 0.193 0.189 0.170 0.054
FROE
R>¹
0.140 0.166 0.191 0.205 0.212 0.183 0.072
FROE
R>`
0.139 0.165 0.190 0.202 0.215 0.183 0.076
FErr
R
0.028 0.034 0.026 0.032 0.062 0.036 0.034***
FErr
R>¹
0.042 0.055 0.057 0.051 0.093 0.059 0.051***
FErr
R>`
0.042 0.045 0.083 0.063 0.122 0.070 0.080***
Ret36 0.636 0.610 0.560 0.539 0.426 0.553 0.210***
Obs. 2336 2365 2373 2389 2398 11861
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 305
`¹ We selected this deﬁnition so that analyst overoptimism results in positive forecast errors. In
most prior studies, forecast errors are deﬁned with the opposite sign (e.g., O’Brien, 1988). Note also
that we report the forecast errors for the next three years. In subsequent tests, we focus on forecast
errors in threeyear ahead ROE forecasts because the longerterm ROEs have the greatest impact on
our »
estimate. The use of ROE rather than earningspershare (EPS) mitigates stock split timing
problems encountered when comparing forecasted and actual EPSs.
for years t!1. FROE
R>G
is the predicted ROE for year t#i based on I/B/E/S
analyst forecasts, and FErr
R>G
is the average forecast error in the year
t#i forecasts. Speciﬁcally, FErr
R>G
is computed by subtracting each ﬁrm’s
predicted year t#i ROE (FROE
R>G
) from the actual reported ROE in period
t#i (ROE
R>G
), and averaging across all ﬁrms. Analyst overoptimism (pessi
mism) relative to future reported earnings results in more positive (negative)
FErr values.`¹ Ret36 is the average threeyear buyandhold return for each
portfolio. Statistical signiﬁcance of the diﬀerence in means between extreme
quintiles is determined using Monte Carlo techniques.
Table 5 conﬁrms several prior ﬁndings, while highlighting the importance of
analyst forecast errors. Firms are included only if they have the ﬁve years of
historical sales data necessary to compute SG. Consistent with prior studies (e.g.,
FF, 1995; Fairﬁeld, 1994; Bernard, 1994), Panel A shows that lower (higher) B/P
ﬁrms have higher (lower) reported ROEs. Both current year ROEs (ROE
R¹
) and
threeyear ahead ROEs (ROE
R>`
) are signiﬁcantly higher for low B/P ﬁrms.
Analysts are also predicting higher proﬁtability for higher P/B ﬁrms: the average
FROE
R>`
is higher (lower) for low (high) B/P ﬁrms. However, a proper investiga
tion of market eﬃciency should focus, not on forecasted or actual ROEs, but on
the pattern of errors in forecasted ROEs (FErr
R>G
).
Consistent with prior studies — e.g., Fried and Givoly (1982), O’Brien (1988)
— Table 5 shows that analysts are, on average, overlyoptimistic: FErr
R>G
is
positive for all quintiles in all three Panels. The magnitude of the bias in
oneyearahead forecasts is comparable to those reported in prior studies (e.g.,
O’Brien, 1988, Table 3). The two and threeyearahead biases are somewhat
higher, reﬂecting the compounding eﬀects of pevious year forecast errors, as well
as our use of the longterm growth rate to estimate threeyearahead FROEs.
More importantly, this table reveals several interesting crosssectional pat
terns in analyst forecast errors. Panel A shows a negative relation between B/P
and FErr
R>`
— analysts are most overoptimistic in low B/P ﬁrms. However, this
relation is not monotonic across the quintiles, and does not hold in one and
twoyearahead forecasts. This result suggests that the B/P eﬀect is only tangen
tially related to FErr.
Panel B shows that SG is positively correlated with FErr
R>`
. While this
ﬁnding is consistent with the LSV mispricing conjecture, the eﬀect is again not
monotonic. Analyst overoptimism increases sharply for the highest SG quintile,
but is otherwise ﬂat. Similarly, buyandhold returns are ﬂat for quintiles Q1
306 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
through Q4, but drop sharply for the top SG quintile. This result is consistent
with Dechow and Sloan (1997), who also ﬁnd that ﬁrm rankings on past growth
measures (both earnings and sales) do not result in a strong systematic return
diﬀerentials in intermediate portfolios. Overall, the evidence suggests SG is
related to analyst forecast errors in the direction predicted by the mispricing
hypothesis, but the relation is nonlinear.
Panel C shows that the OP portfolios are also related to analyst forecast
errors. As predicted by the mispricing hypothesis, analyst forecasts tend to be
more overoptimistic for high OP ﬁrms in all forecast horizons. Similarly, Panel
D shows that analyst forecasts tend to be more overoptimistic for high ¸tg
ﬁrms. Panel D shows that ¸tg has strong predictive power for returns, and
unlike SG, the intermediate portfolio returns to ¸tg rankings are monotonic.
Taken together, Panels A through D show that all four variables appear to have
some predictive power for crosssectional diﬀerences in longterm analyst fore
cast errors.
To evaluate the robustness of these relations over time, Table 6 reports the
result of 15 annual crosssectional regressions of realized analyst forecast errors
on each of these four ﬁrm characteristics (year t"1977—1991). The dependent
variable for each regression is FErr
R>`
. The independent variables are
SG, B/P, OP, or ¸tg. To facilitate interpretation of these results and to reduce
the eﬀect of outliers, the independent variables are expressed in terms of their
percentile ranks. To compute its percentile rank, each variable is sorted as of the
end of June in year t and assigned to percentiles.
Table 6 shows that the relation between these four variables and the sub
sequent analyst forecast error is robust over time. Model 1 shows that low (high)
B/Ps are generally associated with ex post analyst optimism (pessimism). Model
2 shows that high (low) past sales growth (SG) is positively associated with
excessive optimism (pessimism). Model 3 shows that when »
is much higher
(lower) than »
, analysts tend to be too optimistic (pessimistic). Finally,
Model 4 shows that high ¸tg ﬁrms tend to have overly optimistic forecasts. The
sign of the estimated coeﬃcient is correct in 55 out of 60 individual cases.
Newey—West (1987) tstatistics based on timeseries variations in the annual
estimates (bottom row) indicate statistical signiﬁcant at the 1% level for all four
models.
Because the independent variables are all expressed in terms of percentile
ranks, we can compare their estimated coeﬃcients. Based on the last row in
Table 6, the top B/P ranked percentile ﬁrms have forecasted ROE errors that
are 2.5% lower than those of the bottom percentile B/P ﬁrms. Similarly, the
diﬀerence in forecasted ROE errors for the top and bottom SG percentiles is
4.3%. The diﬀerence between top and bottom percentile OP ﬁrms is 7.0%, while
the diﬀerence between top and bottom percentile Ltg ﬁrms is 7.3%. Since the
typical ﬁrm earns an ROE of 13%, these diﬀerences appear substantial and
economically signiﬁcant.
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 307
Table 6
The relation between analyst forecast errors and ex ante ﬁrm characteristics
This table reports the results of 15 (t"1976—1990) crosssectional regressions of realized analyst
forecast errors on four ﬁrm characteristics. Forecast errors (the dependent variable) are computed by
subtracting actual returnsonequity (ROEs) in period t#2 from predicted t#2 ROEs obtained
from I/B/E/S consensus earning forecasts available prior to June 30 of year t. The independent
variables are computed as follows. SG is the ﬁveyear percentage growth in sales from period t!6 to
t!1. B/P is the booktomarket ratio, where B is the reported book value per share for the ﬁscal
year ended in year t!1 and P is the stock price at the end of June in year t. OP is a measure of
analyst optimism derived using Edward—Bell—Ohlson (EBO) fundamental value measures. Speciﬁ
cally, OP"(»
!»
)/"»
", where »
is an EBO value derived using current I/B/E/S analyst
consensus forecasts and »
is a similar EBO value measure derived using historical ROEs. ¸tg is the
consensus longterm earnings growth forecast from I/B/E/S as of May of year t. RK(.) is a percentile
rank operator. To compute percentile ranks, the independent variables are sorted as of the end of
June each year and assigned to percentiles. Numbers in the All Years row represent timeseries
means and Newey—West (1987) tstatistics based on timeseries variation in the annual estimates.
***, **, and * signify statistical signiﬁcance at the 1%, 5% and 10% levels respectively.
Model 1 Model 2 Model 3 Model 4
Year t RK(B/P) RK(SG) RK(OP) RK(¸tg) Obs.
Coeﬀ. 1976 !0.020 0.011 0.049 0.049 206
tstat !1.20 0.67 3.01*** 3.02***
Coeﬀ. 1977 0.005 0.016 0.035 0.017 173
tstat 0.25 0.77 1.71** 0.83
Coeﬀ. 1978 !0.007 0.001 !0.016 !0.023 261
tstat !0.45 0.64 !0.92 !1.27
Coeﬀ. 1979 !0.009 0.029 0.015 0.022 387
tstat !0.68 2.12*** 1.10 1.56*
Coeﬀ. 1980 !0.050 0.036 0.061 0.081 631
tstat !3.28*** 2.37*** 4.02*** 5.40***
Coeﬀ. 1981 !0.130 0.088 0.140 0.150 684
tstat !8.09*** 5.27*** 8.71*** 9.48***
Coeﬀ. 1982 !0.013 0.060 0.003 0.115 694
tstat !0.82 3.87*** 0.20 7.66***
Coeﬀ. 1983 !0.074 0.043 0.143 0.128 684
tstat !4.00*** 2.25*** 7.97*** 7.11***
Coeﬀ. 1984 !0.029 0.046 0.169 0.086 729
tstat !1.59* 2.56*** 11.19*** 4.96***
Coeﬀ. 1985 !0.000 0.048 0.068 0.100 691
tstat !0.016 2.93*** 4.24*** 6.40***
Coeﬀ. 1986 0.002 0.067 0.085 0.081 669
tstat 0.104 3.96*** 5.13*** 4.89***
Coeﬀ. 1987 !0.033 0.040 0.113 0.090 678
tstat !1.84** 2.19*** 6.50*** 5.09***
Coeﬀ. 1988 !0.008 0.077 0.069 0.064 678
tstat !0.447 4.37*** 4.03*** 3.68***
Coeﬀ. 1989 0.001 0.047 0.029 0.059 756
tstat 0.028 2.47*** 1.54* 3.19***
Coeﬀ. 1990 !0.010 0.036 0.081 0.078 779
tstat !0.647 2.20** 5.09*** 4.92***
Mean All !0.025 0.043 0.070 0.073 15 years
tstat Years !2.69*** 7.00*** 5.01*** 6.57***
Table 7
Predicting analyst forecast errors using multiple ﬁrm characteristics
This table reports the results of 15 (t"1976—1990) multiple crosssectional regressions of realized
analyst forecast errors on four ﬁrm characteristics. Forecast errors for each ﬁrm (the dependent
variable) are computed by subtracting actual returnsonequity (ROEs) in period t#2 from
predicted t#2 ROEs obtained from I/B/E/S consensus earning forecasts. To reduce the eﬀect of
outliers, the top and bottom 1% forecasted error each year are omitted. The independent variables
are computed as follows. SG is ﬁve year percentage growth in sales from period t!6 to t!1. B/P is
the booktomarket ratio, where B is the reported book value per share for the ﬁscal year ended in
year t!1 and P is the stock price at the end of June in year t. OP"(»
!»
)/"»
"; where »
is an
EBO value derived using current I/B/E/S analysts consensus forecasts and »
is a similar EBO value
measure derived using historical ROEs. ¸tg is the consensus longterm earnings growth forecast
from I/B/E/S. RK(.) is a percentile rank operator. To compute percentile ranks, the independent
variables are sorted as of the end of June each year, and assigned to percentiles. Numbers in the All
Years row represent timeseries means and Newey—West (1987) tstatistics based on timeseries
variation in the annual estimates. ***, ** and * signify onetailed statistical signiﬁcance at the 1%,
5% and 10% levels, respectively.
Year t RK(SG) RK(B/P) RK(OP) RK(¸tg) R` Fstatistic Obs.
Coeﬀ. 1977 0.010 !0.013 0.032 0.031 0.061 3.31** 206
tstat 0.58 !0.74 1.62 1.62
coeﬀ. 1978 0.020 0.015 0.042 !0.010 0.023 1.00 173
tstat 0.92 0.66 1.52 !0.35
Coeﬀ. 1979 !0.002 !0.019 !0.012 !0.020 0.010 0.62 261
tstat !0.09 !0.89 !0.49 !0.91
Coeﬀ. 1980 0.029 0.009 0.006 0.018 0.017 1.64 387
tstat 2.01 0.57 0.37 1.03
Coeﬀ. 1981 0.014 !0.020 0.023 0.065 0.055 9.13*** 631
tstat 0.88 !1.13 1.26 3.98
Coeﬀ. 1982 0.040 !0.073 0.076 0.096 0.195 41.2*** 684
tstat 2.36 !4.25 4.43 5.64
Coeﬀ. 1983 0.031 0.042 !0.030 0.132 0.097 18.5*** 694
tstat 1.86 2.48 !1.87 7.59
Coeﬀ. 1984 0.048 0.033 0.119 0.089 0.117 22.5*** 684
tstat 2.47 1.36 5.92 3.56
Coeﬀ. 1985 0.079 0.002 0.183 0.000 0.150 31.9*** 729
tstat 4.46 0.10 9.62 0.00
Coeﬀ. 1986 0.042 0.067 0.042 0.109 0.087 16.4*** 691
tstat 2.52 3.70 2.51 5.73
Coeﬀ. 1987 0.062 0.044 0.068 0.055 0.075 13.5*** 669
tstat 3.46 2.44 3.82 2.71
Coeﬀ. 1988 0.020 0.007 0.096 0.056 0.075 13.7*** 678
tstat 1.03 0.35 5.23 2.52
Coeﬀ. 1989 0.064 0.023 0.056 0.025 0.049 8.66*** 678
tstat 3.33 1.22 3.06 1.17
Coeﬀ. 1990 0.030 0.029 0.012 0.056 0.019 3.66*** 756
tstat 1.46 1.45 0.62 2.43
Coeﬀ. 1991 0.014 0.041 0.065 0.072 0.054 11.0*** 779
tstat 0.80 2.24 3.82 3.56
Mean All 0.035 0.010 0.051 0.050 0.074 15 years
tstat Years 5.94*** 1.16 3.55*** 4.26***
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 309
5.5. Predictability of analyst forecast errors
We now assess the amount of crosssectional variation in forecast errors that
is explained by combining our four variables. Table 7 reports the results of
annual multiple regressions of these forecast errors on ranked percentiles of
B/P, SG, OP, and ¸tg. With each variable conditioned on the others, OP, SG
and ¸tg provide greater explanatory power for forecast errors than B/P. In fact,
the NeweyWest (1987) timeseries tstatistics in the last row show that after
controlling for the other variables, B/P oﬀers little incremental contribution to
the model.
The R` for the annual regressions ranges from a low of 1% in 1979 to a high of
19.5% in 1982. The average R` is 7.4%, and the annual Fstatistics indicate
signiﬁcance in 12 out of 15 years. Clearly, a modest, but consistent, portion of
the error in the consensus I/B/E/S forecast is predictable each year. Although
the predictable portion is not large, the consistency of the coeﬃcients suggests
a potential trading strategy. We examine this possibility in the next section.
5.6. Proﬁting from the forecast error
To exploit the predictable component of the I/B/E/S forecast error, we ﬁrst
estimate annual crosssectional regressions of the form presented in Table 7.
Speciﬁcally, we regress forecast errors realized in year t!1 on percentile ranks
of SG, B/P, OP and ¸tg from year t!4. From these annual regressions, we
derive estimated coeﬃcient weights for each variable. We then apply these
estimated coeﬃcients to each ﬁrm’s period t!1 B/P, SG, OP, and ¸tg variables
to compute a predicted forecast error. Speciﬁcally, the predicted forecast error
for ﬁrm i in portfolio formation year t is:
PE
PPGR
"L#K
¹
RK(SG
GR¹
)#K
`
RK(BP
GR¹
)#K
`
RK(OP
GR¹
)
#K
"
RK(¸tg
GR¹
). (4)
The parameters L, K
¹
, K
`
, K
`
and K
"
are estimated from rolling crosssectional
regressions based on year t!4 information and actual year t!1 earnings.
RK(.) is the percentile rank operator. Large positive (negative) values of PErr
correspond to excessively optimism (pessimism) forecasts. Since three past years
of data are necessary to estimate PErr, the sample period for this test is
1979—1992.
Table 8 presents returns to a PErr strategy and compares the results to
returns from other investment strategies. To construct this table, we regressed
the one and threeyearahead buyandhold returns for our sample ﬁrmyears
on scaled decile ranks of ME, B/P, »
/P, and PErr. Following Bernard and
Thomas (1990) and Dechow and Sloan (1997), we use scaled decile rankings for
310 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
all independent variables. For each calendar year, the independent variables are
assigned in descending order to deciles, and then scaled so that they range from
zero (for the lowest decile) to one (for the highest decile). This approach allows
the regression coeﬃcients to be interpreted as estimates of the return to a zero
investment portfolio with a long position in the stocks in the highest decile and
a short position in the stocks of the lowest decile.
Table 8 shows that a decilebased PErr strategy yields approximately 4%
over the next 12months, and 27.7% over the next 36months. Both results are
statistically signiﬁcant at the 1% level. Clearly, PErr has signiﬁcant predictive
power for one and threeyearahead returns. Over the next 12 months, the PErr
strategy performs about as well as »
/P and B/P. Over the next 36 months,
the PErr strategy outperforms the B/P strategy, but underperforms the »
/P
strategy.
Models 5 and 6 in Table 8 evaluate the incremental contribution »
/P and
PErr controlling for B/P and ME. These results show that »
/P has signiﬁcant
incremental power to predict crosssectional returns in both one and three
yearahead regressions. Model 5 in Panel B shows that over 36 months, »
/P is
the most important variable in explaining crosssectional returns. This evidence
extends the results in Table 4 by illustrating that »
/P has incremental predic
tive power controlling for both ME and B/P. In addition, Model 6 in both
panels shows that the PErr strategy enhances the predictive power of »
/P, even
controlling for ME and B/P.
Fig. 2 presents a comparison of the cumulative monthly returns produced by
four alternative trading strategies: a B/P strategy, a PErr strategy, a »
/P
strategy, and a combined strategy. Returns to the B/P and »
/P strategies are
based on buying ﬁrms in the top quintile and selling ﬁrms in the bottom quintile
each year. The results are the same as those reported in Fig. 1. For the PErr
strategy, cumulative returns are the average returns from selling ﬁrms in the top
quintile (high PErr ﬁrms) and buying ﬁrms in the bottom quintile (low PErr
ﬁrms). For the combined strategy, we buy (sell) ﬁrms that are simultaneously in
the top (bottom) »
/P quintile and the bottom (top) PErr quintile. Fig. 2 shows
that the highest returns are produced by the combined strategy, indicating that
the PErrbased strategy has incremental explanatory power to »
/P. Indeed,
over a 36month holding period, this combined strategy yields a cumulative
return of 45.5%.
Table 9 reports the yearbyyear results of implementing each strategy. This
test has less statistical power to detect abnormal returns, but provides a better
picture of the robustness of each strategy over time. The results show that none
of the strategies perform particularly well over oneyear holding periods. How
ever, over threeyear holding periods, PErr, »
/P and the combined strategy all
outperform the B/P strategy.
In the early years (pre1982), the »
/P strategy does better than the combined
strategy. Recall from Table 7 that during these years our prediction model
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 311
Table 8
Relative importance of ex ante ﬁrm characteristics in returns prediction
This table presents estimated coeﬃcients from regressions of one and threeyearahead stock
returns on various ex ante ﬁrm characteristics. The sample consists of 11,861 ﬁrmyears between
1976 and 1992 for ﬁrms with all data available. Dependent variables are the one and threeyear
ahead buyandhold returns, including dividends and any liquidating distributions. The return
cumulation period begins at the end of June in each year t. The independent variables are market
value of equity (ME), booktoprice (BP), valuetoprice (»
/P), and the predicted analyst forecast
error (PErr). Market value of equity (ME) and stock price (P) are as of June 30 of each year t. Book
value (B) is book equity per share in calendar year t!1. »
is a fundamental value estimate derived
using current I/B/E/S consensus forecasts available prior to June 30 of year t. PErr is the predicted
error in the year t consensus forecast of year t#2 ROEs. For ﬁrm i and period t, the predicted
forecast error is
PErr
GR
"L#K
¹
RK(SG
GR¹
)#K
`
RK(BP
GR¹
)#K
`
RK(OP
GR¹
)#K
"
RK(¸tg
GR¹
).
This predicted error is estimated using information available prior to June of year t. Speciﬁcally, we
require each ﬁrm’s ﬁveyear past sales growth (SG), markettobook ratio (BP), longterm consensus
earnings growth forecast (¸tg), and an optimism measure (OP"(»
!»
)/"»
"), where »
is similar
to »
, but derived using historical earnings rather than analyst forecasts. RK(.) is a percentile rank
operator. The parameters L, K
¹
K
`
K
`
, and K
"
are estimated from rolling crosssectional regressions
based on year t!4 information and year t!1 reported earnings. Large positive (negative) values of
PErr correspond to predictions of excessive overoptimism (pessimism). All independent variables
are assigned in descending order to deciles, and then scaled so that they range from zero (for the
lowest decile) to one (for the highest decile). ***, **, * Denote signiﬁcance at the 1%, 5% and 10%
levels, respectively, using a twotailed ttest.
Panel A: Oneyearahead returns
Model Intercept BP ME »
/P PErr Adj. R` (%)
1 0.151*** 0.051*** — — — 0.12
2 0.186*** — !0.019 — — 0.01
3 0.155*** — — 0.042*** — 0.09
4 0.196*** — — — !0.040*** 0.07
5 0.147*** 0.039*** !0.013 0.031** — 0.15
6 0.176*** 0.029* !0.023 0.030** !0.035** 0.19
Panel B: Threeyearahead returns
1 0.468*** 0.168*** — — — 0.37
2 0.538*** — 0.026 — — 0.00
3 0.365*** — — 0.370*** — 1.83
4 0.688*** — — — !0.277*** 1.03
5 0.341*** 0.051 0.013 0.352*** — 1.83
6 0.539*** !0.029 !0.053* 0.343*** !0.241*** 2.47
performed poorly. Thus, the PErr strategy appears to add noise without adding
additional predictive power. However, since 1982, the combined strategy out
performed »
/P every year. The 36month return to the combined strategy is
consistently positive through up and down markets. Yet the strategy is not
312 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
Fig. 2. A comparison of cumulative (buyandhold) returns from alternative trading strategies. This
ﬁgure shows the cumulative returns from several alternative trading strategies. B is book equity per
share in calendar year t!1. P is price per share at the end of June 30 of year t. PErr is the predicted
error in consensus analyst forecasts for year t#2 returnonequity (ROE), estimated as of June 30 of
year t. »
is a fundamental value estimate based on the consensus analyst forecast. Each year,
portfolios are formed at the end of June by sorting ﬁrms into quintiles on the basis of B/P, PErr, and
»
/P. For the B/P and »
/P based strategies, this graph depicts the cumulative buyandhold returns
produced by buying ﬁrms in the top quintile and selling ﬁrms in the bottom quinitle at the beginning
of July, and maintaining these investments until the end of the indicated month. The PErrbased
strategy is simialr, except ﬁrms in the top quintile are sold and ﬁrms in the bottom quintile are
purchased. For the combined PErr and »
P strategy, ﬁrms are included in the long (short) portfolio
if they are simultaneously in the top »
/P quintile and the bottom PErr quintile. The sample period
is 1979—1991 (year t"1970 to 1991).
without risk, particularly over a oneyear horizon. In 1981, the combined
strategy lost 66%, due to large losses in a few stocks.
6. Summary
In this study, we operationalized an analystbased residual income model and
used the resulting valuetoprice (»
/P) ratio to examine issues related to market
eﬃciency and the predictability of crosssectional stock returns. Our results
showthat »
/P is a reliable predictor of crosssectional returns, particularly over
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 313
Table 9
Yearbyyear returns to various trading strategies
This table reports the cumulative oneyear and threeyear buyandhold returns from alternative
trading strategies. B is book equity per share in calendar year t!1. P is price of the stock at the end
of June 30 in year t. PErr is the predicted error in consensus analyst forecasts of year t#2
returnonequity. »
is the EBO fundamental value measure based on the consensus analyst forecast.
Each year, portfolios are formed at the end of June by sorting ﬁrms into quintiles on the basis of B/P,
PErr, and »
/P. For B/P and »
/Pbased strategies (Panels A and C, respectively), the table values
represent the average cumulative equalweighted returns produced by buying ﬁrms in the top
quintile and selling ﬁrms in the bottom quintile at the beginning of July of each year, and
maintaining these investments for either 12 or 36 months. The PErrbased strategy (Panel B) is
similar, except ﬁrms in the top quintile are sold and ﬁrms in the bottom quintile are bought. For the
combined PErr and »
/P strategy (Panel D), ﬁrms are included in the long (short) portfolio if they
are simultaneously in the top »
/P quintile and the bottom PErr quintile. Numbers in the Mean row
represent timeseries means of the annual returns. Reported tstatistics are based on timeseries
variations in the annual means, with Newey—West (1987) correction for serial correlation. Number
of ﬁrms indicates the highest and lowest number of trading positions taken per year, where a trading
position maybe either a long position or a short position. ***, ** and * signify onetailed statistical
signiﬁcance at the 1%, 5% and 10% levels respectively.
Panel A Panel B Panel C Panel D
B/P PErr »
/P Combined
(high—low) (low—high) (high—low) PErr and »
/P
Year t 1year 3year 1year 3year 1year 3year 1year 3year
78 !0.159 0.043 !0.247 !0.294 !0.102 0.325 !0.239 0.171
79 0.091 0.675 0.062 0.338 0.052 0.553 0.022 0.456
80 0.244 0.431 !0.155 !0.092 0.244 0.447 0.077 0.295
81 !0.158 0.286 !0.322 0.529 !0.130 0.393 !0.659 0.909
82 0.234 0.670 0.273 0.745 0.274 1.127 0.343 1.204
83 0.118 0.199 0.261 0.349 0.25 0.589 0.363 0.411
84 !0.105 0.044 0.129 0.264 0.075 !0.002 0.133 0.171
85 0.063 0.176 0.071 0.397 !0.083 0.017 !0.104 0.328
86 0.05 0.052 0.103 0.183 0.072 0.141 0.113 0.228
87 0.199 0.083 0.124 0.221 0.057 0.250 0.190 0.410
88 !0.176 !0.110 !0.019 0.188 !0.062 0.136 !0.038 0.499
89 0.008 0.189 0.074 0.322 0.099 0.214 0.109 0.402
90 0.091 — 0.066 — !0.027 — 0.005 —
91 0.147 — !0.038 — !0.184 — !0.090 —
Mean 0.046 0.228 0.027 0.263 0.038 0.349 0.016 0.457
tstat 1.23 3.32** 0.62 3.55*** 1.02 4.06*** 0.07 5.40***
Number
of ﬁrms
164 to
442
164 to
392
164 to
442
164 to
393
164 to
442
164 to
393
48 to
149
48 to
149
314 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
longer horizons. Over the next 12months, the predictive power of »
/P is
comparable to that of B/P. However, over the next 36months, »
/P has much
stronger predictive power than B/P. This ability to predict longterm returns is
not attributable to B/P, ﬁrm size, or beta.
Because of its importance in estimating »
, we also investigate the reliability
of longterm I/B/E/S consensus earnings forecasts. We ﬁnd that crosssectional
errors in threeyearahead consensus forecast are predictable. Speciﬁcally, we
ﬁnd some evidence that analysts tend to be more overlyoptimistic in ﬁrms
with higher past sales growth and P/B ratios. In addition, we ﬁnd
stronger evidence of overoptimism in ﬁrms with higher forecasted earnings
growth (¸tg) and higher forecasted ROEs relative to current ROEs (OP). Com
bining these variables in a prediction model, we develop an estimate of the
prediction error in longterm forecasts (PErr), and show this estimate has
predictive power for crosssectional returns. Moreover, we show this predictive
power is incremental to a »
/P strategy, and a combined strategy yields the
highest returns.
Our evidence suggests that ﬁrm value estimates based on a residual income
model may be a useful starting point for predicting crosssectional stock returns.
Much recent research has focused accountingbased ratios that exhibit predic
tive power for stock returns. The B/P ratio, in particular, has received signiﬁcant
attention. Our results suggest that superior return prediction may result from
adopting a more complete valuation approach.
Our implementation of the residual income model is simple, and leaves much
room for improvement. While we focus on an analystbased valuation model,
future work may focus on alternative mechanical models of earnings prediction
(e.g., Dechow et al., 1997). Future studies may also lead to reﬁnements in other
key parameters of the model, including forecasted dividend payout ratios and
crosssectional variations in discount rates. We hope that our ﬁndings will
encourage further research along these lines.
Our ﬁnding that prices converge to value estimates gradually over longer
horizons (beyond 12months) is puzzling. The eﬀect may be due, in part, to the
conservative nature of our tests. In a recent replication of our results, Herzberg
(1998) shows a strong partial price correction in the ﬁrst month after the strategy
is implementable (see Exhibit 12). However, we implement this strategy with
a 5 to 6 week lag — we use IBES forecasts publicly available by the third week of
May and form portfolios as of June 30th. Our ﬁrstyear results therefore do not
capture the shortterm proﬁts from the ﬁrst 6 weeks.
This explanation suggests our ﬁrst year returns should be higher, but it does
not explain why returns remain high in years two and three. One explanation is
that the price convergence to value is a much slower process than prior evidence
suggests. This possibility raises interesting questions about the eﬃciency of the
market, and in particular, about the process by which information about
longterm fundamentals is impounded in price. The bias in longterm analyst
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 315
forecasts is not realized until two or three years into the future, and analysts
appear to revise their longterm forecasts only gradually over time. Our evidence
on the predictability of longterm forecast errors in consensus forecasts is consis
tent with this longterm mispricing hypothesis. However, it is diﬃcult to under
stand why arbitrage forces do not eliminate this pricing anomaly more quickly.
Alternatively, »
/P may be yet another proxy for crosssectional risk diﬀer
ences. Our tests control for two obvious sources of potential risk: the B/P ratio
and ﬁrm size. We ﬁnd that high »
/P ﬁrms generally have lower market betas, so
sensitivity to overall market movements is an unlikely explanation for their
higher subsequent returns. In addition, unlike returns to a B/P strategy, returns
to a »
/P strategy exhibits a pattern of lower shortterm returns and higher
longterm returns. This pattern is diﬃcult to reconcile with a risk explanation.
Despite these concerns, we acknowledge that high »
/P ﬁrms may still be riskier
than low »
/P ﬁrms in some other, as yet unidentiﬁed, dimension. We leave this
question to future research.
7. For Further Reading
The following references are also of interest to the reader: Dechow et al. (1998)
and Fama and French (1955).
Acknowledgements
We thank the late Victor Bernard, whose many insights helped to bring the
residual income valuation model to life. We also thank Jeﬀ Abarbanell, Jim
Bodurtha, Larry Brown, John Core, Kent Daniel, Tom Dyckman, Ken French,
S.P. Kothari (Editor), Bruce Lehman, Pat O’Brien, Jay Shanken (the referee),
Richard Sloan, Bhaskaran Swaminathan, an anonymous referee, and workshop
participants at Cornell University, Dartmouth College, Georgetown University,
Harvard University, the University of Minnesota, Ohio State University, the
University of Oregon, the University of Rochester, and Yale University for helpful
suggestions. James Myers provided expert research assistance. Steve Merritt,
a Michigan MBA student, deserves credit for ﬁrst identifying an apparent market
anomaly with this trading rule during a class exercise. Earnings forecasts used in
this paper are provided by I/B/E/S. We gratefully acknowledge the ﬁnancial
support of the QGroup and the KPMG Peat Marwick Foundation (Lee).
Appendix A. Using I/B/E/S forecasts to derive future ROE estimates
Our implementation of the Edwards—Bell—Ohlson (EBO) formula requires
three future ROE forecasts [FROE
, FROE
R>¹
and FROE
R>`
]. We derive these
316 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
future ROEs from I/B/E/S consensus EPS estimates. Since yearend book values
are dependent on current year ROEs, we use a sequential process to estimate
future ROEs. The steps in the process are listed below. Year t refers to the year of
portfolio formation.
Step 1: Estimating FROE
R
and B
. We require that all sample ﬁrms have
a oneyearahead I/B/E/S consensus EPS forecast [F½1]. Forecasted ROE for
year t is then computed as the year t consensus forecast, divided by the average
book value per share during year t!1. Use of the average, rather than yearend,
book value reduces the chance of an extremely low denominator. We then use
FROE
R
and the dividend payout ratio (k) to derive the ending book value for
year t. Notationally, we have:
FROE
R
"F½1/[(B
R¹
#B
R`
)/2],
B
R
"B
R¹
[1#FROE
R
(1!k)].
Step 2: Estimating FROE
R>¹
and B
R>¹
. We also require that all sample ﬁrms
have a twoyearahead consensus forecast [F½2]. We then compute FROE
R>¹
and B
R>¹
analogously:
FROE
R>¹
"F½2/[(B
R
#B
R
!1)/2], B
R>¹
"B
R
[1#FROE
R>¹
(1!k)].
Step 3: Estimating FROE
R>`
and B
R>`
. Where a longterm earnings growth
estimate [¸tg] is available, we compute FROE
R>`
and B
R>`
as follows:
FROE
R>`
"[F½2(1#¸tg)]/[(B
R>¹
#B
R
)/2],
B
R>`
"B
R>¹
[1#FROE
R>`
(1!k)].
Where ¸tg is not available, we use FROE
R>¹
to proxy for FROE
R>`
.
References
Abarbanell, J., 1991. Do analysts’ earnings forecasts incorporate information in prior stock price
changes?. Journal of Accounting and Economics 14, 147—165.
Abarbanell, J., Bernard, V., 1992. Tests of analysts’ overreaction/ underreaction to earnings informa
tion as an explanation for anomalous stock price behavior. Journal of Finance 47, 1181—1207.
Abarbanell, J., Bernard, V., 1995. Is the U.S. stock market myopic? Working paper, University of
Michigan, January.
Ball, R., Kothari, S.P., Shanken, J., 1995. Problems in measuring portfolio performance: an applica
tion of contrarian investment strategies. Journal of Financial Economics 38, 79—107.
Barber, B.M., Lyon, J.D., 1997. Detecting longrun abnormal stock returns: The empirical power
and speciﬁcation of test statistics. Journal of Financial Economics 43, 341—372.
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 317
Basu, S., 1977. Investment performance of common stocks in relation to their price earnings ratios:
a test of the eﬃcient market hypothesis. Journal of Finance 32, 663—682.
Bernard, V.L., 1994. Accountingbased valuation methods, determinants of booktomarket ratios,
and implications for ﬁnancial statement analysis. Working paper, University of Michigan,
January.
Brown, L.D., Han, J.C.Y., Keon, E., Jr., Quinn, W.H., 1995. Predicting analysts’ earnings surprise.
Journal of Investing, forthcoming.
Brown, L.D., Griﬃn, P.A., Hagerman, R.L., Zmijewski, M.E., 1987a. Security analyst superiority
relative to univariate timeseries models in forecasting quarterly earnings. Journal of Accounting
and Economics 9, 61—87.
Brown, L.D., Richardson, G.D., Schwager, S.J., 1987b. An information interpretation of ﬁnancial
analyst superiority in forecasting earnings. Journal of Accounting Research 25, 49—67.
Chan, L.K.C., Hamao, Y., Lakonishok, J., 1991. Fundamentals and stock returns in Japan. Journal
of Finance 46, 1739—1764.
Daniel, K., Mande, V., 1994. Business cycle variation in earnings forecasts and common
stock returns. Working paper, University of Chicago and University of Nebraska at Omaha,
September.
Davis, J.L., 1994. The crosssection of realized stock returns: the preCompustat evidence. Journal of
Finance 48, 1579—1593.
DeBondt, W.F.M., 1993. Betting on trends: intuitive forecasts of ﬁnancial risk and return. Interna
tional Journal of Forecasting 9, 355—371.
Dechow, P.M., Sloan, R.G., 1997. Returns to contrarian investment: tests of the naive expectations
hypothesis. Journal of Financial Economics 43, 3—27.
Dechow, P., Hutton, A., Sloan, R., 1998. An empirical assessment of the residual income valuation
model. Working paper, Forthcoming, Journal of Accounting Economics, 26.
Edwards, E., Bell, P., 1961. The Theory and Measurement of Business Income. University of
California Press, Berkeley, CA.
Fairﬁeld, P., 1994. P/E, P/B and the present value of future dividends. Financial Analysts Journal
July/Aug., 23—31.
Fama, E.F., French, K.R., 1992. The crosssection of expected stock returns. Journal of Finance 47,
427—465.
Fama, E.F., French, K.R., 1997. Industry costs of equity. Journal of Financial Economics 43,
153—193.
Fama, E.F., French, K.R., 1995. Size and booktomarket factors in earnings and returns. Journal of
Finance, forthcoming.
Feltham, G.A., Ohlson, J.A., 1995. Valuation and clean surplus accounting for operating and
ﬁnancial activities. Contemporary Accounting Research 11, 689—731.
Frankel, R., Lee, C.M.C., 1998. Accounting diversity and international valuation. Working paper,
University of Michigan and Cornell University, January.
Fried, D., Givoly, D., 1982. Financial analysts’ forecasts of earnings: a better surrogate for market
expectations. Journal of Accounting and Economics 4, 85—107.
Herzberg, M.M., 1998. Implementing EBO/EVA analysis in stock selection. The Journal of Invest
ing 45—53.
Jaﬀe, J., Keim, D.B., Westerﬁeld, R., 1989. Earnings yields, market values, and stock returns. Journal
of Finance 44, 135—148.
Kothari, S.P., Shanken, J., Sloan, R., 1995. Another look at the crosssection of expected returns.
Journal of Finance 50, 185—224.
Kothari, S.P., Warner, J.B., 1997. Measuring longhorizon security price performance. Journal of
Financial Economics 43, 301—340.
Lakonishok, J., Shleifer, A., Vishny, R.W., 1994. Contrarian investment, extrapolation, and risk.
Journal of Finance 49, 1541—1578.
318 R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
LaPorta, R., 1996. Expectations and the crosssection of stock returns. Journal of Finance 51,
1715—1742.
LaPorta, R., Lakonishok, J., Shleifer, A., Vishny, R., 1997. Good news for value stocks: further
evidence on market eﬃciency. Journal of Finance 52, 859—874.
Lee, C.M.C., 1996. Measuring wealth CA Magazine April, 32—37.
Lee, C.M.C., Myers, J., Swaminathan, B., 1998. What is the intrinsic value of the Dow? Journal of
Finance, forthcoming.
Lyon, J.D., Barber, B.M., Tsai, C.L. 1998. Improved methods for tests of longrun abnormal stock
returns. Journal of Finance 53, forthcoming.
Lehman, B., 1993. Earnings, dividend policy, and present value relations: building blocks of
dividend policy invariant cash ﬂows. Review of Quantitative Finance and Accounting 3,
263—282.
Litzenberger, R.H., Ramaswamy, K., 1979. The eﬀect of personal taxes and dividends on capital
asset prices. Journal of Financial Economics 7, 163—195.
O’Brien, P.C., 1988. Analysts’ forecasts as earnings expectations. Journal of Accounting and
Economics 10, 53—83.
Ohlson, J.A., 1990. A synthesis of security valuation theory and the role of dividends, cash ﬂows, and
earnings. Contemporary Accounting Research 6, 648—676.
Ohlson, J.A., 1991. The theory of value and earnings, and an introduction to the BallBrown
analysis. Contemporary Accounting Research 7, 1—19.
Ohlson, J.A., 1995. Earnings, Book Values, and Dividends in Security Valuation. Contemporary
Accounting Research 11, 661—687.
Ou, J.A., Penman, S.H., 1994. Financial statement analysis and the evaluation of booktomarket
ratios. Working paper, Santa Clara University and the University of California at Berkeley, May.
Peasnell, K., 1982. Some formal connections between economic values and yields and accounting
numbers. Journal of Business Finance and Accounting 361—381.
Penman, S.H., 1995. A synthesis of equity valuation techniques and terminal value calculations for
the dividend discount model. Working paper, U.C. Berkeley, February.
Penman, S.H., Sougiannas, T., 1998. A comparison of dividend, cash ﬂow, and earnings approaches
to equity valuation. Working paper, University of California at Berkeley, and University of
Illinois at UrbanaChampaign, October.
Preinreich, G., 1938. Annual survey of economic theory: the theory of depreciation. Econometrica 6,
219—241.
Stober, T.L., 1992. Summary ﬁnancial statement measures and analysts’ forecasts of earnings.
Journal of Accounting and Economics 15, 347—372.
Tversky, A., Kahneman, D., 1984. Judgment under uncertainty: heuristics and biases. Science 185,
1124—1131.
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 319
284
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
earnings forecasts and examine its usefulness in predicting crosssectional stock returns in the U.S. Speciﬁcally, we use I/B/E/S consensus earnings forecasts to proxy for market expectations of future earnings. We then use the resulting estimate of ﬁrm fundamental value (» ) to investigate issues related to market eﬃciency and the predictability of crosssectional stock returns. We ﬁnd that » estimates based on I/B/E/S consensus forecasts are highly correlated with contemporaneous stock prices. In recent years, » explains more than 70% of the crosssectional variation in stock prices. Moreover, the valuetoprice ratio (» /P) is a good predictor of crosssectional returns, particularly over longer time horizons. In 12month horizons, the » /P ratio predicts crosssectional returns as well as the booktomarket ratio (B/P). However, over two or three year periods, buyandhold returns from » /P strategies are more than twice those from B/P strategies. Speciﬁcally, we ﬁnd that higher » /P ﬁrms tend to earn higher longterm returns. This result is not due to diﬀerences in market betas, ﬁrm size, or the B/P ratio. Because of its importance in estimating » , we also investigate the reliability of longterm I/B/E/S consensus earnings forecasts. We ﬁnd that crosssectional errors in the threeyearahead consensus forecast are predictable. Speciﬁcally, we ﬁnd some evidence that analysts tend to be more overlyoptimistic in ﬁrms with higher past sales growth (SG) and higher P/B ratios. In addition, we ﬁnd stronger evidence of overoptimism in ﬁrms with higher forecasted earnings growth (¸tg) and higher forecasted ROEs relative to current ROEs (OP). Combining these variables in a prediction model, we develop an estimate of the prediction error in longterm forecasts (PErr), and show this estimate has predictive power for crosssectional returns. Finally, we show the predictive power of PErr is incremental to a » /P strategy. During our sample period (1979—1991), a zerocash investment strategy involving ﬁrms that are simultaneously in the top quintile of » /P and the bottom quintile of PErr yields cumulative buyandhold returns of more than 45% over 36 months. The threeyear buyandhold strategy results in positive returns in both up and down markets. This eﬀect is not explained by market beta, ﬁrm size, or the B/P ratio. Our results contribute to the emerging literature on the residual income model in several ways. First, our analystbased approach complements Penman and Sougiannas (1998), which uses ex post reported earnings. Second, we provide evidence on the reliability of I/B/E/S consensus forecasts for valuation, as well as a method for correcting predictable forecast errors. To our knowledge, this is the ﬁrst study to develop a prediction model for longrun analyst forecast errors, and to trade proﬁtably on that prediction. Finally, we show that returns
Several studies show analyst forecast errors diﬀer for ﬁrms with certain characteristics, suggesting a relation between analyst forecast errors and various market pricing anomalies (e.g., Dechow and Sloan, 1997; Daniel and Mande, 1994; LaPorta, 1996; LaPorta et al., 1997). Other studies show
R. Frankel, C.M.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319
285
to a » /P strategy are not due to standard risk proxies, and that the strategy can be further improved by incorporating analyst forecast errors. Our ﬁndings are also related to the ﬁnance literature on the predictability of stock returns. Much recent research has focused on accountingbased ratios that exhibit predictive power for stock returns. The B/P ratio, in particular, has been elevated to celebrity status by studies such as Fama and French (1992). Fama and French suggest B/P is a proxy for a ﬁrm’s distress risk. However, little progress has been made in identifying the exact nature of this risk. Our results suggest that rather than attempting to produce a better risk proxy, superior return prediction may result from adopting a more complete valuation approach. In sum, empirical studies involving equity valuation encounter two potential problems: (1) the use of overly restrictive models of intrinsic value, and (2) the use of biased proxies as model imputs. Our research design features a more robust valuation model than simple marketmultiples, as well as a technique for improving on analysts’ earnings forecasts. Our empirical ﬁndings suggest both will lead to better predictions of crosssectional stock returns. The remainder of this paper is organized as follows. In the next section, we present the accountingbased valuation model and describe its most salient features. In Section 3, we discuss the estimation procedures used to implement this model. Section 4 contains a discussion of the data and sample description. Section 5 reports the empirical results, and Section 6 concludes with a summary of our ﬁndings and their implications.
2. The residual income model The valuation method we use in this study is a discounted residual income approach sometimes referred to as the Edwards—Bell—Ohlson (EBO) valuation technique. Independent derivations of this valuation model have surfaced periodically throughout the accounting, ﬁnance and economics literature since the 1930s. In this section, we present the basic residual income equation and brieﬂy develop the intuition behind the model.
analysts may not use all available information when formulating their forecasts (e.g., Abarbanell, 1991; Abarbanell and Bernard, 1992; Stober, 1992). However, none of these studies develop a prediction model for analyst errors. Brown et al. (1995) do develop a prediction model for analyst errors, but their investment horizon is only onequarterahead and the details of their model are proprietary. The term Edwards—Bell—Ohlson, or EBO, was coined by Bernard (1994). Theoretical development of this valuation method is found in Ohlson (1990, 1995), Lehman (1993), and Feltham and Ohlson (1995). Earlier treatments can be found in Preinreich (1938), Edwards and Bell (1961), and Peasnell (1982). For a simple guide to implementing this technique, see Lee (1996).
In other words. the model provides a framework for analyzing the relation between accounting numbers and ﬁrm value. R>G Note that this equation is identical to a dividend discount model. E [.] is expectation based on information R R available at time t. deﬁned as present value of future discounted cash ﬂows not captured by the current book value. Notationally. However. Frankel. C. However. and r is the cost of equity capital based on the information set at time t. It therefore relies on the same theory and is subject to the same theoretical limitations as the dividend discount model. E (D ) R R>G . then Eq. will be worth only their current book value. (1) can be rewritten as the reported book value. Dividing both sides of Eq. This deﬁnition assumes a ﬂat termstructure of discount rates. plus an inﬁnite sum of discounted residual income: E [NI !(r B )] R R>G R>G\ »H"B # R R (1#r )G G E [(ROE !r ) B ] R R>G R>G\ . Eq. "B # (2) R (1#r )G G where B is the book value at time t. r is the cost of R>G equity capital and ROE is the aftertax return on book equity for period t#i. »H is the stock’s fundamental value at time t. but expresses ﬁrm value in terms of accounting numbers. and » "B .M. as long as a ﬁrm’s earnings and book value are forecasted in a manner consistent with clean surplus accounting. Lee / Journal of Accounting and Economics 25 (1998) 283–319 A stock’s fundamental value is typically deﬁned as the present value of its expected future dividends based on all currently available information. a typical ﬁrm’s ROE should be close to its . In a competitive equilibrium.C. E (D ) is the R R R>G expected future dividends for period t#i conditional on information available at time t. NI is the Net Income for period t#i. then the present value of future residual income is zero. If a ﬁrm earns future accounting income at a rate exactly equal to its cost of equity capital. (2) oﬀers a natural interpretation for the pricetobook ratio. R R ﬁrms that neither create nor destroy wealth relative to their accountingbased shareholders’ equity. ﬁrms whose expected ROEs are higher (lower) than r will have values greater (lesser) than their book values.286 R. (1) »H. R (1#r )G G In this deﬁnition. (2) by B . Eq. (2) shows that equity value can be split into two components — an accounting measure of the capital invested (B ). we can express P/B in terms of a ﬁrm’s future abnormal R ROEs. It is easy to show that. and a measure of the present R value of future residual income. If the market price approximates future discounted cash ﬂows.
1994. 1994. Later. . this equation properly impounds a ﬁrm’s expected future proﬁtability into its equity value estimate. ﬁrms’ reported ROE may diﬀer from their costs of equity in competitive equilibrium due to accounting and risk factors. 1977. Except for Dechow et al. we use analyst forecasts of future earnings to directly examine whether the market fully incorporates current information in establishing prices. Jaﬀe et al. For example. Fairﬁeld. FF. ﬁrms expected to earn above (below) normal ROEs in the future should trade at higher (lower) pricetobook ratios. (1997) and Dechow et al. 1994). and the typical pricetobook ratio should be close to 1. (1997) examine the empirical properties of the model under alternative speciﬁcations. Current literature shows that a number of ad hoc variables such as cash ﬂow yield (Chan et al. Several recent studies evaluate this model’s ability to explain stock prices. earnings yield (Basu. Davis.C. we show later (in Table 1) that average P/Bs are above 1. These yield measures have often been interpreted as risk proxies. Stated another way. Eq. not that it fully incorporates all available information when doing so. LSV. conservative. One of the objectives in this paper is to use the R EBO model to derive a more precise estimate of »H. Lee / Journal of Accounting and Economics 25 (1998) 283–319 287 cost of equity capital (ROE"r ). In practice. More recently. 1979) have predictive power for crosssectional returns. (1997). Penman and Sougiannas (1998) implement variations of the model using ex post realizations of earnings to proxy for ex ante expectations. (1998) operationalize the model for the 30 stocks in the Dow Jones Industrial Average and examine timeseries properties of the model. on average. This association is consistent with the empirical ﬁnding that low (high) B/P ﬁrms have higher (lower) future ROEs (e. 1991. However. Eq. Moreover.g. neither book value nor earnings. and average ROEs are somewhat above ﬁrms’ costs of capital.. C. both Francis et al. Bernard.. these studies do not examine the predictive power of the model for crosssectional stock returns in the US. is R suﬃcient to capture »H .M. (2) shows that future earnings performance should be closely linked to current B/P ratios.. Frankel and Lee (1998) employ the model in an international context and ﬁnd that » has high explanatory power for prices in 21 countries. Frankel. These observations are consistent with the fact that accounting systems that are. Lee et al. 1989) and dividend yield (Litzenberger and Ramaswamy.R. 1995. and examine whether R a more complete valuation model yields superior power to predict risk adjusted returns. However. the inverse relation between future earnings performance and current B/P ratios should not be interpreted as proof of market eﬃciency — it shows that the market considers future proﬁtability when formulating prices. (2) suggests that these market multiples may also work because their accounting component reﬂects (imperfectly) some dimension of »H. 1994).
For example. as well as a value metric based on consensus I/B/E/S analyst forecasts (» ). there is little consensus on how this discount rate should be determined. we discuss the speciﬁcs of the model estimation procedure.M. For this study. We use this measure as a proxy for normal earnings when reported earnings are negative. Two alternatives. Frankel.C. The dividend payout ratio is the percentage of net income paid out in the form of dividends each year.g. In this section.. Six percent reﬂects the average longrun returnonassets (see Table 1. based on ex ante information. Dividend payout ratio (k). The FF discount rates are based on a onefactor and a threefactor risk model. in conjunction B . to derive forecasted book values: R B "B #NI !d "B #(1!k)NI R> R R> R> R R> "[1#(1!k)ROE ]B . last column). we use three diﬀerent approaches — a constant discount rate. In theory. is used in conjunction with the clean surplus relation (CSR) to derive future book values. we can write B "[1#(1!k)ROE ][1#(1!k)ROE ]B . The ﬁrst three parameters’ roles are readily seen in Eq. equivalently. we also constrain k to be between 0 and 100%. or (2) use analysts’ earnings forecasts (e. (2). As explained later.288 R. and R future ROEs. Cost of equity capital (r ). R The four parameters needed for the estimation are: the cost of equity capital (r ). (2) presents a simple procedure for estimating a ﬁrm’s intrinsic value (» ). the dividend payout ratio (k). however. We obtain a ﬁrmspeciﬁc estimate of k by dividing the common stock dividends paid in the most recent year (Compustat Item 21) by net income before extraordinary items (Compustat Item 237). Abarbanell and Bernard (1995) use earnings forecasts from ValueLine]. future ROE forecasts (FROEs). current book value (B ). r should be ﬁrmspeciﬁc. We use both methods and derive a value metric based on historical earnings (» ). Lee / Journal of Accounting and Economics 25 (1998) 283–319 3. Model estimation procedures Eq. and a dividend payout R ratio (k). forecasting future earnings). . For ﬁrms with negative earnings (approximately 11% of our sample). k. C. In practice. The last input. all future book values can be expressed as functions of B . we divide dividends by six percent of total assets to derive an estimated payout ratio. and two industrybased discount rates derived by FF (1997). This variable is used. R> R Analogously. R> R> R> R Future ROEs. reﬂecting the premium demanded by equity investors to invest in a ﬁrm or project of comparable risk. are: (1) use prior period earnings (or ROEs). The most important and diﬃcult task in the EBO valuation exercise is forecasting future ROEs (or.
For example. Various alternative approaches have appeared in the literature. (1997) feature various permutations for the terminal value. (1994) show that.b). suggesting that the current period ROE is a reasonable starting point for estimating future ROEs. The 12period version had slightly lower correlation with stock prices and similar predictive power for returns. the explicit forecast period must be ﬁnite. In theory. However. (2) expresses ﬁrm value in terms of an inﬁnite series. markets are myopic.66. Prior studies show that analyst earnings forecasts are superior to timeseries forecasts (e. We also estimated a 12period expansion of the formula in which ROEs are reverted back to the industry median. 1995).g. This limitation necessitates a terminal value estimate — that is. For example. One approach is to estimate the terminal value by ﬁrst expanding Eq. we estimate three forms of » : R (FROE !r ) (FROE !r ) R R »"B # K B# B.S. (2) to ¹ terms. In this study. They also provide evidence that future forecasted ROEs may not be fully impounded in current prices. The resulting value estimate therefore depends critically on the particular earnings forecast used in the terminal value. and then taking the next term in the expansion as a perpetuity. 1988. an estimate of the value of the ﬁrm based on residual income earned after the explicit forecasting period.. However.C. O’Brien. we take a simple approach using a shorthorizon earnings forecasts of up to three years. R R R R (1#r ) (1#r )r (3. if the explicit forecast period ends after ¹ periods. (1998) and Dechow et al. ¹ should be set large enough for ﬁrms to reach their competitive equilibrium..M. both Lee et al. Frankel. but for practical purposes. the predictive superiority of an analystbased value metric (» ) over a historicalbased value metric (» ) is an open empirical question. Eq. 2 (1#r )2r This procedure is mathematically equivalent to a ¹period discounted dividend model in which year ¹#1 earnings is treated as a perpetuity (see Penman. Forecast horizons and terminal value estimation. Therefore. they examine whether markets underprice longrun earnings relative to nearterm earnings. . in large samples. Lee / Journal of Accounting and Economics 25 (1998) 283–319 289 Fairﬁeld et al. the correlation between current year ROEs and next year’s ROEs is around 0. C. The use of I/B/E/S data should result in a more precise proxy for market expectations of earnings.1) Abarbanell and Bernard (1995) use ValueLine forecasts to estimate a similar EBO valuation equation in addressing the question of whether U. our ability to forecast future ROEs diminishes quickly over time. 1987a. the terminal value is: !r ) (ROE 2> B . Brown et al.R. Under the null hypothesis of market eﬃciency. and forecasting errors are compounded in longer expansions.
ROE "NI /[(B #B )/2].1) represents a twoperiod expansion of the residual income model with the forecasted ROE for the current year (FROE ) assumed to be earned in R perpetuity. but we use a twoyearahead forecasted ROE (FROE )in the perpetuity. We further constrain our sample to ﬁrms with ﬁscalyearends between June and December. and B is total common shareholders’ equity from year t (Compustat Item 60). we derive future ROEs and book values from I/B/E/S consensus forecasts using a sequential procedure described in the Appendix. C. and NASDAQ return ﬁles from the Center for Research in Security Prices (CRSP) and (b) a merged Compustat annual industrial ﬁle. and R\ R\ R\ DI» ) and have the necessary CRSP stock prices and shares outstanding data R\ (for ﬁscal year end t!1. inclusively. R> Eq. 4. (3. In theory. . B . The results are also similar for a sample consisting of just December yearend ﬁrms. we require ﬁrms to have a oneyearahead and a twoyearsahead earningspershare (EPS) forecast from I/B/E/S. Eq.3) is a threeperiod model. and the end of June in year t). R R R R\ to proxy for all future ROEs — i. R> (1#r )r (3. Furthermore. we use the annual average to avoid situations where an unusually low book value in year t1 inﬂates forecasted ROEs. We require ﬁrms to meet the Compustat data requirements (for B . and preferred dividends (Compustat Item 237). Data and sample description The original sample of ﬁrms consists of all domestic nonﬁnancial companies in the intersection of (a) the NYSE. Lee / Journal of Accounting and Economics 25 (1998) 283–319 (FROE !r ) (FROE !r ) R R> »"B # K B# B . net of R extraordinary items. taxes. this requirement limits our sample period to 1975—93. including PST. the model calls for beginningofyear book values.. To estimate » . R R R R> (1#r ) (1#r )r (FROE !r ) (FROE !r ) R R> »"B # K B# B R R R R> (1#r ) (1#r ) (FROE !r ) R> # B . (3. NI is earnings to common shareholders in year t.M. Because I/B/E/S began operations in 1975. Because we use I/B/E/S forecasts issued in May.e. NI . we use the return on average equity.2) (3. Frankel. The righthand side of each equation consists of ex ante observables. this constraint ensures that forecasted earnings correspond to the correct ﬁscal year. To R estimate » .3) Eq.C. (3.290 R. Using » for the entire Fama and French (1992) sample over the period of 1962—1993 yields similar results.2) also represents a twoperiod expansion of the model. AMEX. full coverage and research ﬁles. Similarly. However. we substitute ROE for all the FROEs in the R above equations.
The Appendix explains the procedure we followed to derive future ROE forecasts when all three variables are not in the May I/B/E/S report. » /P and » /P ratios and poor market liquidity (that is.C.R.126 observations (approximately 5%). Using median rather than mean forecasts is unlikely to aﬀect results because the distribution of forecasted growth is quite symmetric. These procedures eliminate 1075 ﬁrmyears. leaving a ﬁnal sample of 18. (3. (3.162 ﬁrmyears. In addition. in the sense that all the portfolios are constructed using ﬁrm characteristics that are observable at the time of portfolio formation. Taken together. we match accounting data for all ﬁscal year ends in the calendar year t!1 to returns on portfolios formed at the end of June of year t. We eliminate such ﬁrms by considering only ﬁrms with ROEs or FROEs of less than 100% and dividend payout ratios of less than 100%. and the market equity on June 30 of year t to measure its size. rather than ﬁscal year end. In estimating » . Since these monthly reports are widely available soon after each computer run. they cannot be included in equalweighted portfolios without incurring disproportionally large trading costs). or earnings. the May statistics are in the public domain well before our portfolio formation date.1). These procedures are similar to FF (1992).2) and (3. We also remove 51 ﬁrms with stock prices of under $1 as of the end of June in year t. These ﬁrms have unstable B/P.3). leading to unreasonable ROE or k estimates. Our valuation formula uses three pieces of I/B/E/S data: earningspershare forecasts oneyearahead (F½1). In estimating Eqs. (1995) for additional evidence on the sensitivity of contrarian returns to the removal of stocks with prices under $1. EPS forecasts twoyearsahead (F½2). Lee / Journal of Accounting and Economics 25 (1998) 283–319 291 To ensure that accounting variables are known before returns are computed. See Ball et al. Speciﬁcally. This mean estimate is determined from analyst forecasts on ﬁle with I/B/E/S as of the Thursday after the third Friday of each month. analysts reported just F½1 and F½2. because ROEs for these ﬁrms cannot be interpreted in economic terms. . the strategies we examine are tradable. They report that this diﬀerence has little eﬀect on their return tests. some ﬁrms have extremely low book values. Further. Between 1975 and 1979. we remove ﬁrms with negative book values. except their B/P ratios are based on December market prices. we use the I/B/E/S mean (also called consensus) forecast from the May statistical period of year t.M. These common sense ﬁlters ensure the subsequent results are not driven by outliers. our ﬁlters eliminated 1. but after 1980. Frankel. we allow a minimum gap of six months between the ﬁscalyearend and the portfolio formation date. C. and a ﬁveyear longterm growth rate (¸tg). We use a ﬁrm’s market equity at its ﬁscalyearend to compute its booktomarket and valuetomarket ratios. most ﬁrms also had ¸tg information.
22 0.23 0.94 1.64 1.96 1.19 0. P is the stock price as of June 30 in year t. these results illustrate the stability of the key model parameters over our sample period.06).06 0. k is computed as common stock dividends divided by (total assets .22 0. The average dividend payout ratio ranges from a high of 35% in 1976 and 1978 to a low of 19% in 1993.82 2.52 1.11 0.10 0.74 14.43. Table 1 Summary statistics by year Table values represent annual.0.292 R.05 0.13 Avg.12 0.30 1.08 0. Averages reported in the bottom row represent timeseries means of the annual statistics.M.52 1.53 1.21 0.43 Avg. Year t!1 is the year from which the accounting data are obtained. The average P/B ratio is 2.04 0. computed as common stock dividends divided by earnings to common shareholders.71 1.36 10.87 Avg.25 0.26 20.02 1. ME 1168 1264 863 740 875 921 693 1049 781 1002 1269 1387 1282 1423 1506 1591 1579 1605 1167 Avg. The average book value per share over the period is $16.11 2. Finally.83 22.99 2.59 1.73 12.07 0. Frankel. B is the year t!1 reported book value per share.32 0.18. Lee / Journal of Accounting and Economics 25 (1998) 283–319 5. For ﬁrms with negative earnings.54 2.66 15. Year t No.15 0.61 1.34 2.33 2.51 22.32 0.07 0. The average returnonequity ranges between 8 and 18%.60 1.04 0.69 22.65 12.14 0.77 22.45 1. Taking the inverse of the B/P ratio [shown as 1/Avg(B/P)] reduces the average P/B ratio to 1.34 1.51 2.17 0.83 13.82 1.69 2.17 16.38 2.96 2. 1/(Avg.09 0.67 1.22 0. C.08 0.06 0.B/P) is the inverse of the equallyweighted average B/P ratio for each year.05 0.11 0. B/P) 1.57 11.44 18.27 0.11 0.81 20. Collectively.33 0.35 0.C.27 Avg.06 0.08 0.69 1.05 1.33 0.04 0.06 0.63 2. Empirical results Table 1 reports annual summary statistics for the total sample.13 0.05 0.87. Table 1 reports the returnonasset ratio [Avg ROA].08 0. however.23 0.07 0.32 0.54 1.44 1.13 0.17 12. which averages 6%.15 1.16 0. equallyweighted average statistics for the sample ﬁrms.25 0. ME is the market value of equity as of June 30 of year t in millions of dollars.44 15.13 1.40 1. P/B 2.09 0.08 0.32 0. B 21.18 1/(Avg.81 1.06 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 All years . ROA 0.35 0. k 0.18 0. ROE 0. ROE is the return on equity for year t!1 computed as net income for year t!1 divided by the year t!1 average book equity.97 12. this ratio is inﬂated by the presence of low book value ﬁrms in the sample.16 0. ROA is the return on total assets for year t!1.12 0. ﬁrm 361 312 535 675 718 812 920 952 1174 1130 1146 1213 1228 1298 1306 1352 1423 1607 1 8162 Avg.27 0. k is the dividend payout ratio.05 0.16 0.
These six measures reﬂect the three diﬀerent empirical estimates of value discussed earlier (Eqs.74 0.77 0.76 0.68 0.80 0.1).83 0.80 0.60 0.70 0. Lee / Journal of Accounting and Economics 25 (1998) 283–319 293 5.89 0.72 0.79 0.82 0. The discount rate used is a threefactor industryspeciﬁc costofequity (Fama and French. The historical EBO value measures use current year returnonequity (ROE) to proxy for future ROEs.81 0.74 0.56 0.73 0.85 0.89 0.76 0.45 0.89 0.82 0.82 0.48 0.69 Analyst based EBO value measures 0.69 0.49 0.66 0.84 0.87 0.68 0.68 0. 1997).71 0.88 0.70 0.70 0.87 0.53 0. FF Threefactor ¹"1 Year t Obs.82 0.R.79 0.64 0.86 0.76 0.87 0. Each set of nine value measures diﬀers only in the assumed the number of forecasting periods (¹"1.74 0.82 0.83 0.74 0.74 0.76 0.85 0.87 0.66 0.87 0.63 0.60 . All years represents the timeseries mean of annual crosssectional correlations.87 0.68 0.69 0. estimated using historical earnings and analyst forecasts.70 0.73 0. book equity per share in calendar year t!1 (B). Correlation with stock prices Table 2 reports crosssectional Spearman rank correlation coeﬃcients between stock prices and either book value (B) or one of six EBO value metrics.1.74 0.73 0.74 0.79 0.77 0.M.60 0.72 0.70 0.45 0.75 0. and two sets of three fundamental value metrics computed using the Edwards—Bell—Ohlson (EBO) formula.69 0. (3.79 0.68 0.88 0.68 0.67 0.43 0.77 0.83 0.77 0.73 0.69 0. 2.88 0. The analyst based EBO value measures use consensus I/B/E/S forecasts to proxy for future ROEs.64 0.70 0.82 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 All years 361 312 535 675 718 812 920 952 1174 1130 1146 1213 1228 1298 1306 1352 1423 1607 1 8162 0.79 0.72 0.81 0.79 0. The discount rates used are industry speciﬁc costofequity Table 2 Annual crosssectional correlation of stock prices to book equity and EBO value measures Table values represent crosssectional Spearman correlation coeﬃcients between the stock price on June 30 of year t.76 0. or 3).71 0.68 0.53 0.71 0.80 0. C.78 0.65 0.61 0.C.81 0.67 0.77 0.80 0.68 0.61 0. Frankel.82 0.72 0.71 0.78 0.72 0.73 0.56 0.63 0.68 0.87 0. B ¹"2 ¹"3 FF Threefactor ¹"1 ¹"2 ¹"3 Historical EBO value measures 0.73 0.72 0.73 0.86 0.72 0.64 0.81 0.82 0.86 0.64 0. (3.68 0.3)).72 0.54 0.2) and (3.54 0.65 0.78 0.
we use » to denote the fundamental value computed from historical ROEs. In sum. book value (B) had an average correlation with price of 0. we use the risk premium for each industry reported in Fama and French (1997).C.M. . In estimating these models. 1982. Compared to B.49% (for Real Estate). a three factor industryspeciﬁc discount rate. using a threeperiod model for » yields similar results. suggesting that book equity explains around 36% of the crosssectional variation in prices. Speciﬁcally. C. Table 7. The threefactor model is estimated using mimicking portfolios for ﬁrm size and markettobook ratios.0646 per year — the average annualized 30day tbill rate over our sample period.b.. The superiority of analysts over simple randomwalk models in forecasting earnings is well documented (Fried and Givoly. We use » to denote the fundamental value computed using the mean analyst forecast. Our ﬁndings suggest analyst forecasts also better reﬂect the market expectations of earnings implicit in the EBO model. » explains around twothirds of the crosssectional variation in prices. 1997). 1988).13 We also used constant interest rates of 11%. The risk premiums we used are computed from 5year rolling regressions by industry. » explains around 49% of the crosssectional variation in prices. and a forecast horizon of three periods (Eq. Over our sample period.294 R. and a forecast horizon of twoperiod (Eq. However. Evidently. probably due to rapid decay in the precision of the ROE forecasts over time. Table 2 suggests that using all variations of the estimated » metric is unnecessary. values based on historical ROEs contain important valuerelevant information not captured by B. See Fama and French (1997) for more details. Abarbanell and Bernard (1995) also ﬁnd that allowing for intertemporal and ﬁrmspeciﬁc variations in r had little C eﬀect on their results. We ﬁnd varying the discount rate had little eﬀect on our results. The threefactor industry costofequity ranges between 8. and 13% as well as an industryspeciﬁc singlefactor model. Therefore. ﬁrms are grouped into 48 industry classes. Brown et al. We used a twoperiod model for » because of concerns for the accuracy of the timeseries earnings model beyond two years. O’Brien. analysts’ earnings forecasts contain more valuerelevant information than is reﬂected in historical ROEs. Lee / Journal of Accounting and Economics 25 (1998) 283–319 based on a threefactor model (Fama and French. Table 2 also shows that the crosssectional correlation with price increases when EBO value is estimated using analyst forecasts. (3. (3. Increasing ¹ from 1 to 3 produces slightly better correlations. Increasing ¹ from 1 to 3 produces slightly weaker correlations.60.3)).23% (for Alcoholic Beverages) and 16. 12%. Frankel. but varying the discount rate again has little eﬀect. 1987a. plus a constant riskless rate of 0. Over our sample period. for the remainder of this paper.2)). each of the six value measures displays a higher average correlation with stock price. a threefactor industryspeciﬁc discount rate.
Speciﬁcally.M. we use pvalues calculated from the simulated empirical distribution of mean Q5—Q1 returns. The statistical signiﬁcance of this diﬀerence is assessed using a Monte Carlo simulation technique similar to those discussed in Barber and Lyon (1997). ME.549.C. the use of pvalues calculated from the simulated empirical distribution avoids the skewness bias discussed in the literature. To determine statistical signiﬁcance.R. and Lyon et al. and adjusts for serial correlations in returns induced by overlapping holding periods.377 for Ret12. and 12. and » /P. Table 3 reports the average B/P. For each portfolio.2. and 36 months (Ret36). both reﬂect buyandhold returns over the same time horizon). Ret24 and Ret36. First. and » /P values. Our randomization procedure avoids the three main econometric problems discussed in Lyon et al. Kothari and Warner (1997).385. (1998) and Kothari and Warner (1997). we compute returns for the reference portfolios in exactly the same manner as for the actual portfolios (that is. Finally. B/P. Market beta for each ﬁrm is estimated using an equalweighted market index and each ﬁrm’s monthly returns over the next 36 months. and average buyandhold return over the next 12 months (Ret12). 24 months (Ret24). Second. and applies to all variables except the stock return variables. Lee / Journal of Accounting and Economics 25 (1998) 283–319 295 5. We then compute the mean returns for the Q5—Q1 portfolio. respectively. This procedure generates ﬁve random quintile portfolios each year with the same number of observations as the actual quintile portfolios. the size decile cutoﬀs are based on June 30 prices for all NYSE ﬁrms. 14. We repeat the process until we have obtained 1000 sets of quintile portfolios for each year. we use annual insample ﬁrm size cutoﬀs to form quintiles. we construct unidimensional portfolios based on market value of equity (ME). As a ﬁrst step. our reference portfolios only contain ﬁrms that are available for investing at the same time as our sample ﬁrms. Firm size quintiles are formed in two ways. (1998). The last row in each panel shows the number of ﬁrmyear observations in each portfolio. we form empirical reference distributions by randomly assigning the population of eligible ﬁrms each year into quintile portfolios (without replacement). This avoids the rebalancing bias. The right column of Table 3 reports the diﬀerences in means between the top (Q5) and bottom (Q1) quintiles. C. This table is constructed by sorting all sample ﬁrms into quintiles at the end of June each year. Correlation with future returns: unidimensional analyses The main focus of this study is on the prediction of future returns. . Second. When we require availability of stock returns. First. as well as the average postranking market betas. This avoids the new listing or survivor bias. as in Fama and French (1992). Frankel. Table 3 reports characteristics of quintile portfolios formed on the basis of these ﬁrm characteristics. the number of observations drop to 16.
311 0. **.161 0. !0.98 0.497 9333 All Firms 1230 0.045** 0. ***. Lee / Journal of Accounting and Economics 25 (1998) 283–319 Table 3 Characteristics of quintileportfolios formed by ME. Panel A — Marketequity portfolios (NYSE size quintiles) Q1 Q2 Q3 Q4 Q5 (Low ME) (High ME) ME B/P » /P Beta Ret12 Ret24 Ret36 Obs.001 0.311 0. we form empirical reference distributions by randomly assigning eligible ﬁrms into quintiles each year.493 1 8162 Q5!Q1 Diﬀ. twoyear.233*** !0.338* Panel B — Marketequity portfolios (insample size quintiles) Q1 Q2 Q3 Q4 Q5 (Low ME) (High ME) ME B/P » /P Beta Ret12 Ret24 Ret36 Obs 41 0.10 0.310 0. results represent diﬀerences in means between the top (Q5) and bottom (Q1) quintiles.79 0.503 3632 4983 0.92 1. Each panel reports mean values for individual quintile characteristics.17 0.159 0.65 0.60 0.62 0.42 0.486 3636 277 0. and analyst based EBO valuetoprice (» /P).91 1. Analyst based EBO value (» ) is a fundamental value measure derived using current I/B/E/S consensus analyst predictions of future earnings available prior to June 30 of year t. Frankel.08 0.C.07 0.304 0. Price (P) is the stock price at the end of June in year t.39 1.05 !0.85 1.159 0. ME is the market value of shareholders’ equity as of June 30 of year t. 5% and 10% levels. The statistical signiﬁcance of this diﬀerence is derived using Monte Carlo simulation.87 1.95 0.77 !0.04 1.11 0.066** . bookvaluetoprice (B/P).312 0.159 0. respectively (twotailed).515 0.91 1. and » /P This table reports the characteristics of quintile portfolios formed at the end of June each year by market value of equity (ME). and Ret36.61 0. ME quintiles are based on size cutoﬀs for all NYSE ﬁrms (Panel A) and on insample size cutoﬀs (Panel B).835 232 25 1. threeyear buyandhold return for the portfolio.69 0. C. The sample period is 1977—1992 (t"77 to 92).14 0. * signify that the observed diﬀerence between the extreme quintiles is signiﬁcantly diﬀerent from those of the reference distribution at the 1%.210* !0.M.69 0.158 0. !0.442 2692 157 0.525 3640 All Firms 1230 0.379 0.158 0.296 R.92 1.30 !0.90 1. Ret12.06 0.153 0.285 0.493 1 8162 Q5!Q1 Diﬀ.08 0. 9 1. Results in the All Firms column represent unconditional means.556 1144 59 0.33 0.00 1.17 0.311 0. Q5!Q1 Diﬀ. and Ret36 are the average oneyear. Obs.159 0.319 0. is the number of observations in each quintile and applies to all variables except Ret12.32 !0.73 0.03 0.157 0. expressed in millions.330 0. Speciﬁcally. beta is estimated using monthly returns over the 36 months beginning July of year t.69 0. Ret24.459 3622 117 0.239 0.340 0.487 4761 2293 0.305 0.91 1.66 0. B/P.489 3632 722 0. Ret24.08 0.04 1.87 1. Book value (B) is book equity per share in calendar year t!1.146 0.
151*** 0.60 1345 0.02 0.99 0.42 1434 0. this simulation procedure still has a potential deﬁciency.031*** 0.535 3636 Panel D — » /P portfolios Q1 Q2 (Low » /P) » /P B/P ME Beta Ret12 Ret24 Ret36 Obs 0.351 0. the pvalues based on the simulation may be misleading.75 1377 0.159 0.407 3621 0.493 1 8162 Q5!Q1 Diﬀ.M.369 0.97 0. crosscorrelations in the data may cause the variance of the simulated Q5—Q1 returns to be lower than that of the true Q5—Q1 returns.86 1.154 0.300 0. Panel B shows We thank the referee for his insights on this point.450 3628 0.156 0.54 0.85 1177 1.138 0. Lee / Journal of Accounting and Economics 25 (1998) 283–319 Table 3 (continued) Panel C — Booktoprice (B/P) Portfolios Q1 Q2 Q3 (Low B/P) B/P ME » /P Beta Ret12 Ret24 Ret36 Obs 0.166 0. because we require that ﬁrms be followed by analysts.05 0. the procedure creates portfolios whose covariance is equal to the average covariance across all returns in our sample. small ﬁrms generally outperform large ﬁrms.39 666 1. If the actual correlation structure between portfolios one and ﬁve diﬀers signiﬁcantly from this average covariance.513 3634 297 Q4 Q5 (High B/P) 1.169 0.05 0. larger ﬁrms dominate the sample — the last row of Panel A shows that over 80% of our ﬁrms are larger than the median NYSE ﬁrm.40 0.06 0.558 3643 All Firms 0.333 0. C.24 1641 0. and 36 month periods following portfolio formation.450 3632 Q3 Q4 Q5 All (High » /P) Firms 1. Panels A and B examine the ME eﬀect for our sample.C. we will provide additional corroborating evidence using more traditional statistical procedures (see Tables 8 and 9).311 0.637 3640 0. However.137 0. Over 12.049*** 0.298 0.148 0. .81 1068 1.70 0. 24.251 0.549 3632 Despite these advantages.152*** 0.68 1531 1.186 0.338 0.159 0. To mitigate this problem.306*** 0.93 1.75 1.159 0.91 0.17 0.317 0.69 1230 0.210 0. — 0.332 0.172 0.08 0.493 1 8162 Q5!Q1 Diﬀ.260 !0.87 0.91 1.09 0.08 0.60 812 1.29 0.R.69 1230 1.217 0.03 0.331 3626 0.24 0.01 0. By randomly assigning ﬁrms to quintiles 1 through 5.320 0.93 0. Panel A shows that a smallﬁrm eﬀect exists when NYSE size quintiles are used. In particular.25 365 !0.59 1252 1.491 3632 1.311 0. — !975 0. Frankel.082*** 0.
Panel D shows that » /P portfolios have some similarities with B/P portfolios. Thus. Dechow and Sloan. for example. Over 36 months. these results show that » /P also predicts returns.8% over the next 12 months.M. The lowest » /P quintile ﬁrms earn 13.C. Fig. C. the returns pattern is not monotonic across the quintiles. the »/P strategy outperforms the B/P strategy by a wide margin. .6%. a longposition is taken in the top quintile ﬁrms based on each ratio.9% is statistically signiﬁcant at 1%. The relation between B/P and future returns is generally monotonic across the quintiles. Our ﬁndings suggest that while analyst consensus forecasts provide Kothari et al.6%. The lowest B/P quintile ﬁrms earn an average return of 13. (1995) show annual betas are more highly correlated with crosssectional returns. and a shortposition is taken in the bottom quintile ﬁrms.g. Moreover. we ﬁnd low B/P ﬁrms have higher betas than high B/P ﬁrms. The shortterm prediction results for » /P are slightly weaker than the results for B/P. However. The B/P eﬀect is also seen over longer holding periods. The diﬀerence of 4. In summary. Furthermore. we ﬁnd that » /P is much better at explaining crosssectional prices than B/P. 1997). the ﬁrmsize eﬀect is driven primarily by a small set of the smallest stocks. high » /P ﬁrms signiﬁcantly outperform low » /P ﬁrms. tend to have lower market betas. the spread between the highest and lowest » /P portfolios is 30. This result suggests that the B/P eﬀect is not due to diﬀerences in market risk. To construct this graph. Indeed. we observe a monotonic relation in returns across the quintiles. High B/P betas may also be biased downward due to nonsychroneity. The graph shows that over a 36 month period. Interestingly. Frankel. However.9%. » /P and B/P are positively correlated.7% over the next 12months while the highest B/P quintile ﬁrms earn 18. over 24 and 36 months. returns from the B/P strategy exhibit a seasonal pattern not found in the other two strategies — Fig. This graph reports the diﬀerence in cumulative buyandhold returns between the top and bottom quintiles at monthly intervals over the next three years. large ﬁrms actually outperform small ﬁrms over 24 and 36 month holding periods. although the larger nature of our sample ﬁrms may reduce this problem. like high B/P ﬁrms. 1 shows that the January eﬀect is much more pronounced in the B/P strategy than in the »/P strategy.. Results for » /P are similar to those for » /P. in our sample. (1994). while the highest » /P quintile ﬁrms earn only 16. Lee / Journal of Accounting and Economics 25 (1998) 283–319 that when our ﬁrms are grouped by insample size cutoﬀs. and is comparable in magnitude to other studies reporting the B/P eﬀect using I/B/E/Sconstrained samples (e. » /P is also a better predictor of longterm returns. 1 illustrates the cumulative returns from a 36 month buyandhold strategy involving B/P and » /P.298 R. Panel C conﬁrms a B/P eﬀect for our sample. » /P is not necessarily more useful for predicting returns over 12 month intervals. High » /P ﬁrms. over these longer horizon. As in Lakonishok et al.
C. trading on the basis of these forecasts does not necessarily yield higher shortrun (12 month) returns than trading on B/P. Each year.R. portfolios are formed at the end of June by sorting ﬁrms into quintiles on the basis of B/P and » /P. and maintaining these investments until the end of the indicated month. We examine the extent to which these two factors explain the predictive power of » /P. C. The sample period is 1979—1991 (year t"1979 to 1991). we explore a strategy that seeks to improve on the consensus earnings forecast. In later tests. P is price per share at the end of June 30 of year t.M. B is book equity per share in calendar year t!1. 5.3. Correlation with future returns: bidimensional analyses We now consider how much of the explanatory power of » /P for longterm returns is due to its correlation with ﬁrm size and B/P. this graph depicts the cumulative buyandhold returns produced by buying ﬁrms in the top quintile and selling ﬁrms in the bottom quintile at the beginning of July. Lee / Journal of Accounting and Economics 25 (1998) 283–319 299 Fig. 1. Frankel. » is a fundamental value estimate based on the consensus analyst forecast as of May of year t. a good proxy for market expectations. . This ﬁgure shows the cumulative (buyandhold) returns from B/P and » /P based trading strategies. Cumulative (buyandhold) returns produced by B/P and » /P Trading strategies. Fama and French (1992) show that both ﬁrm size and B/P have predictive power for crosssectional returns. For each investment strategy.
we form empirical reference distributions by randomly assigning eligible ﬁrms into quintiles each year while holding quintile membership in the other variable constant.053* 0. Price (P) is the stock price at the end of June in year t.473 (535) 0. 5% and 10% levels. The number of observations is shown in parentheses.578 (665) 0. Book value (B) is book equity per share in calendar year t!1.388*** 0. Panel A — Average 36month buyandhold returns for portfolios formed on the basis of both ﬁrm size (insample quintiles) and » /P Analystbased EBO valuetomarket (» /P) Q2 Q3 Q4 Q5 (High » /P) All Firms Q5!Q1 Diﬀ. ***.486 (2406) 0. C.493 (12377) Q5!Q1 diﬀ.290*** 0. all sample ﬁrms are sorted (independently) into quintiles according to the variables designated in the panels.329*** Q1 (Low » /P) Size quintiles Q1 (Small ME) Q2 Q3 Q4 R.418 (405) 0.590 (668) 0. 0.489 (508) 0.427 (518) 0.318 (529) 0.486 (447) 0.549 (2538) 0.078* 0. Speciﬁcally.C. **.525 (2670) 0.503 (2536) 0. Lee / Journal of Accounting and Economics 25 (1998) 283–319 Q5 (Large ME) All ﬁrms 0.653 (2392) 0.031 0. respectively (onetailed). Analystbased EBO value (» ) is a fundamental value measure derived using I/B/E/S consensus analyst forecasts to proxy for future earnings.525 (425) 0.300 Table 4 Average 36month buyandhold returns for bidimensional portfolios This table reports the average 36month realized returns for bidimensional portfolios.383 (378) 0.577 (460) 0. On June 30 of year t.350 (370) 0.270*** 0.319 497 0.383 (559) 0. The Q5!Q1 diﬀ. Table values represent mean buyandhold returns for each portfolio over the next 36 months.565 (554) 0.089* .679 (595) 0.504 (447) 0.287 (480) 0. results represent diﬀerences in mean returns between the Q5 and Q1 quintiles. controlling for quintile membership in other variable.433 (498) 0.459 (2373) 0.637 (2490) 0. formed using two ex ante ﬁrm characteristics simultaneously.491 (2495) 0.706 (407) 0. The statistical signiﬁcance of this diﬀerence is derived using a Monte Carlo simulation technique.450 (2470) 0.M. Frankel. The sample period is 1977—1991 (t"77 to 91).050 0.557 (360) 0.331 (2384) 0.496 (542) 0.520 (500) 0. 0. * signify that the observed diﬀerence between the Q5 and Q1 quintiles is signiﬁcantly diﬀerent from those of the reference distribution at the 1%. Results for All Firms represent unconditional means.271*** 0.539 (519) 0.498 (511) 0.
!0.493 (12377) 0.C.516 (344) 0.637 (2490) 0.331 (2384) R.318*** 0.558 (2481) All Firms 0.535 (2550) 0.150*** 0.484 (533) 0.026 0.469*** Q3 Q4 Q5 (High » /P) Q5!Q1 Diﬀ.170* 0.342 (296) 0.407 (2419) 0.442 (328) 0.366 (495) 0.491 (2495) 0.513 (2488) 0.440 (515) 0.589 (866) 0.630 (600) 0.450 (2439) 0.549 (2538) 0.489 (573) 0.566 (453) 0.457 (264) 0.461 (694) 0.M.396 (295) 0.053 301 .588 (510) 0.131 0. Frankel.316 (998) 0.350 (210) 0.280*** 0.415 (333) 0.Panel B — Average 36month buyandhold returns for portfolios formed on the basis of both B/P and » /P Analystbased EBO valuetomarket (» /P) Q2 All Firms 0.634 (269) 0.263 (386) 0.576 (565) 0. Q1 (Low » /P) Booktomarket Q1 (Low B/P) Q2 Q3 Q4 Q5 (high B/P) 0. C.732 (824) 0.450 (2470) !0.468 (592) 0.530 (660) 0. Lee / Journal of Accounting and Economics 25 (1998) 283–319 Q5!Q1 Diﬀ.202*** 0.422 (341) 0.098 0.544 (433) 0.
The proceeds from termination. we assigned the total population of ﬁrms within each B/P quintile into random » /P quintiles each year (without replacement). The right column shows that these diﬀerences are statistically signiﬁcant at the 1% level in each of ﬁve size quintiles. We again assess the statistical signiﬁcance of the diﬀerence between Q1 and Q5 portfolios using a Monte Carlo simulation technique. Looking down each column. In this analysis. we examine future returns to » /P portfolios while controlling for ME and B/P. The diﬀerence in Q5—Q1 returns range from 27. the B/P eﬀect survives in only one » /P quintile. and no longer monotonic after controlling for » /P. Panel A reports portfolio returns for » /P and ﬁrm size (using insample size cutoﬀs). . The simulation results show » /P explains longrun returns within all ﬁve B/P portfolios. Table 4 reports the average realized return to 36month buyandhold strategies for bidimensional portfolios. we independently sort ﬁrms into quintiles based on each partitioning variable as of the end of June each year. we hold quintile membership in the other variable constant while randomizing across the variable of interest. Panel B reports portfolio returns for » /P and B/P. Conversely. Panel A shows that » /P has strong predictive power in all ﬁve size quintiles. a terminating return to the delisting date is computed. Panel B shows the interaction of the B/P and » /P eﬀects. we see that the B/P eﬀect is much weaker. Later. Firms that switch exchanges are traced to their new exchange listings and retained in their original portfolios. large ﬁrms slightly outperform small ﬁrms after controlling for » /P. to create the empirical distribution for » /P in Panel B.8%. For example. Panels A and B suggest that in longer time horizons. When a ﬁrm drops out over our holding period. we examine the relative contribution of each variable in returns prediction using a multiple regression approach. C. Frankel. if any. however.C. we observe a largely monotonic and statistically signiﬁcant relation between » /P and returns. This procedure controls for B/P membership each year as well as any serial correlation in yeartoyear returns.302 R.M. looking across each row.0% to 38. Results are similar when we use NYSEsize quintiles. are equally assigned to surviving ﬁrms in the same portfolio. The bottom row shows that when ﬁrms are divided using insample size cutoﬀs. the predictive power of » /P for future returns is not explained by either B/P or ﬁrm size. however some cells have very few observations. Lee / Journal of Accounting and Economics 25 (1998) 283–319 To address this question. Taken together. To construct this table. The resulting empirical distribution allows us to assess the incremental usefulness of one variable in predicting returns after controlling for the other. Stocks are then assigned to one of 25 portfolios based on their bidimensional ranking.
Therefore. LaPorta et al. Lee / Journal of Accounting and Economics 25 (1998) 283–319 303 5. LaPorta (1996). We extend this literature by examining the power of Ltg to predict errors in longterm analyst forecasts. Past studies show that analysts are more accurate than Similar arguments have been made in Dechow and Sloan (1995). (1997). OP is a measure of analyst optimism derived from EBO fundamental value measures. Speciﬁcally. The use of the consensus longterm earnings growth forecast (¸tg) is motivated by LaPorta (1996) and Dechow and Sloan (1997). this investigation should be helpful in distinguishing between the two hypotheses. we should be able to further improve the ability of » /P to predict returns by incorporating the predictable errors. LSV argue that their ﬁnding is due to investor over optimism (pessimism) in ﬁrms with high (low) past sales growth. as well as a trading strategy that directly exploits the predicted error in analysts. ¸tg was not reported by I/B/E/S until 1981. In the pre1981 period. we investigate the relation between analyst forecast errors and four ex ante ﬁrm characteristics — the booktomarket ratio (B/P). Unique features of our study include the use of a comprehensive valuation model that utilizes earnings forecasts the development of a prediction model for analyst forecast errors. the mispricing hypothesis predicts that high (low) SGs are associated with over optimistic (pessimistic) I/B/E/S forecasts. we implicitly assume that these forecasts are unbiased with respect to public information.4. Daniel and Mande (1994). we deﬁne SG in terms of the percentage growth in sales over the past ﬁve years. including strategies based on the B/P and E/P ratios. Speciﬁcally. the mispricing hypothesis suggests a relation between certain characteristics and the direction of subsequent forecast errors. Thus. The relation between analyst forecast errors and ex ante ﬁrm characteristics In this part of the paper. In using I/B/E/S consensus forecasts to estimate » . Frankel.C. we estimated the longterm growth rate using the growth rate implicit between F½1 and F½2. Dechow and Sloan (1997) show that the Ltg eﬀect accounts for a signiﬁcant portion of the return to contrarian investment strategies. while the risk hypothesis does not. alone and in combination with variables. . Second. First.M.R. LaPorta shows that ﬁrms with higher longterm earnings forecasts (high ¸tg ﬁrms) tend to earn lower subsequent returns. C. (1994)’s [LSV] ﬁnding that ﬁrms with higher (lower) past sales growth earn lower (higher) subsequent returns. Like LSV. analyst consensus longterm earnings growth forecast (¸tg). If this assumption is not true. and a new measure we call OP (for analyst optimism). The use of SG is suggested by Lakonishok et al. past sales growth (SG). OP measures the extent to which equity values based on analyst forecasts deviate from similar valuations based on historical earnings. This investigation has two motivations. the reliability of » depends critically on the quality of the earnings forecast used. we investigate the quality of I/B/E/S consensus forecasts. OP"(» !» )/"» ".
OP measures the extent to which equity values based on analyst forecasts deviate from similar valuations based on historical earnings. and ¸tg.037 0.207 0. ***. in the crosssection.051 0.028 0. The statistical signiﬁcance of this diﬀerence is derived using Monte Carlo simulation. We ﬁnd the same is true for our sample of ﬁrms.170 0. Q5!Q1 Diﬀ.019** !0.183 0. 0.157 0. results represent diﬀerences in means between the top (Q5) and bottom (Q1) quintiles.036 0. ¸tg.025 0.185 !0.151*** 0.070 0. SG.C.617 2375 . ﬁrms are required to have all of the above variables available.846 0.051 0.158 0. * signify that the observed diﬀerence between the extreme quintiles is signiﬁcantly diﬀerent from those of the reference distribution at the 1%.208 0.265 0. ¸tg is the consensus longterm earnings growth forecast from I/B/E/S as of May in year t.202 0.104 0.173 0. 5% and 10% levels. Results for All Firms represent unconditional means.124 0. C. **. Panel A — Booktomarket (B/P) portfolios Q1 (Low B/P) B/P ROE R\ FROE R FROE R> FROE R> FErr R FErr R> FErr R> Ret36 Obs. and P is the stock price at the end of June in year t. Lee / Journal of Accounting and Economics 25 (1998) 283–319 earnings forecasts based on timeseries models.049 0. Speciﬁcally.144 0.553 1 1861 Q5!Q1 Diﬀ. Sample period is 19771991 (t"77 to 91). Frankel.642 0. The forecast error for each observation (FErr ) is computed by subtracting the forecasted (FROE ) from the R>G R>G actual reported ROE in period t#i. If analysts underweight historical proﬁtability benchmarks in making their forecasts.200 0.074 0.107 0. B/P is the booktomarket ratio. SG. Table 5 reports characteristics of quintile portfolios formed at the end of June each year by B/P.149 !0. ROE is the actual reported returnsonequity R\ Table 5 Accounting proﬁtability of quintile portfolios formed by B/P.205 !0.129 0.543 2372 0. OP.171 0. respectively (onetailed).183 0. To be included. we form empirical reference distributions by randomly assigning eligible ﬁrms into quintiles each year.073 0. then OP will be positively correlated with analyst optimism.031 0.M.054 0. Each panel reports mean values for individual quintiles.058 0. Two examples from the behavioral literature that discuss this possibility are Tversky and Kahneman (1984) and DeBondt (1993). However.147 0.113 0.366 0. where » is an EBO value derived using current I/B/E/S consensus forecasts and » is a similar EBO value measure derived using historical ROEs.060 0.108 0.462 2350 Q2 Q3 Q4 Q5 (High B/P) 1. is the number of observations.173 0.039** 0. and OP This table reports the accounting proﬁtability characteristics of quintile portfolios formed at the end of June each year. OP captures analyst optimism relative to past reported ROEs — analysts expect the highest (lowest) OP ﬁrms to experience the most (least) ROE growth.062 0.080 0.280 0. OP captures these deviations.059 0.613 2386 All Firms 0.053 0.066 0. — !0.289 0.131 0. where B is the book value per share for the ﬁscal year ended in year t!1.023*** 0. analyst forecasts that deviate the most from historical earnings may reﬂect an underweighting of historical information.304 R. OP"(» !» )/"» ".455 0. Ret36 is the average threeyear buyandhold return for the portfolio.530 2378 0. ROE is equal to G reported net income in year i divided by the average of year i and i!1 book values. Obs. SG is the ﬁveyear growth rate in sales from period t!6 to t!1.285 0. Speciﬁcally.716 0.
185 0.060 0.322 0.205 0.145 0.083 0.034 0.201 0.147 0.046 0.063 0.176 0.042 0.017 0.046 0.053 0.455 2380 All Firms 1.099*** 0.190 0.045 0.179 0.135 0.129 0.129 0.129 0.192 0.066 0.189 0.087 0.154 0.165 0. Frankel.152 0.122 0.018 0.207 0.002 0.179 0.059 0.193 0.118 0.560 2373 0.126 0.093 0.056*** 0.038 0.052** 0.135 0.R. C.081 0.591 2372 0.051*** 0.539 2389 .C.646 0.192 0.032 0.610 2365 0. — !0.183 0. — 0.052 0. Lee / Journal of Accounting and Economics 25 (1998) 283–319 Panel B — Sales growth (SG) portfolios Q1 (Low SG) SG ROE R\ FROE R FROE R> FROE R> FErrt FErr R> FErr R> Ret36 Obs.030 0.23 0.189 0.070 0.141 0.053 0.183 0.207 0.215 0.183 0.105 0.125 0.059 0.193 0.036 0.053 0.185 0.032*** 0.554 2365 Q2 Q3 Q4 Q5 (High OP) 5.054 0.059 0.046 0.070 0.147 0.052*** !0.027 0.174 0.202 0.042 0.076*** 0.186 0.067 0.022 0.553 1 1861 Q5!Q1 Diﬀ.626 2372 0.036 0.426 2398 All Firms 0.028 0.081 0.140 0.212 0.553 1 1861 305 Q5!Q1 Diﬀ.156 0.050 0.072 0.636 2336 Q2 Q3 Q4 Q5 (High ¸tg) 0.051 0.553 1 1861 Q5!Q1 Diﬀ.177 0.034 0.122 0.637 2365 Q2 Q3 Q4 Q5 (High SG) 9.025 0.194 0.105 0.155 0.643 2372 0.126 0. !0.162 0.076 0.178 0.059 0.054 0.170 0.055 0.162 0.283 0.170 0.210*** 0.153 0.166 0.148 0.036 0.120 0.451 0.176 0. !0.174 0.169 0.384 2380 All Firms 2.125 0.057 0.183 0.114 0.170 0.191 0.062 0.051 0.092 0.253*** 0.055 0.161 0.225 0.031 0.070 0. — 0.024 0.192 0.139 0.199 0.148 0.067 0.014 0.132 0. 0.052 0.026 0.033*** 0.034*** 0.143 0.179 0.331 0.057 0.787 0.183 0.074 0.166 0.183 0.556 2372 Panel C — Analyst optimism (OP) portfolios Q1 (Low OP) OP ROE R\ FROE R FROE R> FROE R> FErr R FErr R> FErr R> Ret36 Obs.202 0.558 2372 1.062 0.M.521 2372 Panel D — Longterm growth forecast (¸tg) portfolios Q1 (Low ¸tg) ¸tg ROE R\ FROE R FROE R> FROE R> FErr R FErr R> FErr R> Ret36 Obs.306 0.155 0.194 0.041 0.021 0.016 0.051 0.080*** 0.060 0.
C. we focus on forecast errors in threeyear ahead ROE forecasts because the longerterm ROEs have the greatest impact on our » estimate. The use of ROE rather than earningspershare (EPS) mitigates stock split timing problems encountered when comparing forecasted and actual EPSs. but on the pattern of errors in forecasted ROEs (FErr ). a proper investigaR> tion of market eﬃciency should focus.C. FROE is the predicted ROE for year t#i based on I/B/E/S R>G analyst forecasts. while highlighting the importance of analyst forecast errors. O’Brien (1988) — Table 5 shows that analysts are. on average. FF. R> Analysts are also predicting higher proﬁtability for higher P/B ﬁrms: the average FROE is higher (lower) for low (high) B/P ﬁrms. 1994). Note also that we report the forecast errors for the next three years. This result suggests that the B/P eﬀect is only tangentially related to FErr. buyandhold returns are ﬂat for quintiles Q1 We selected this deﬁnition so that analyst overoptimism results in positive forecast errors. 1994. Table 3). In subsequent tests. and averaging across all ﬁrms.and twoyearahead forecasts. Panel A shows that lower (higher) B/P ﬁrms have higher (lower) reported ROEs. Panel B shows that SG is positively correlated with FErr . Table 5 conﬁrms several prior ﬁndings. While this R> ﬁnding is consistent with the LSV mispricing conjecture. The magnitude of the bias in oneyearahead forecasts is comparable to those reported in prior studies (e.g. this table reveals several interesting crosssectional patterns in analyst forecast errors. O’Brien. However. Both current year ROEs (ROE ) and R\ threeyear ahead ROEs (ROE ) are signiﬁcantly higher for low B/P ﬁrms. . Firms are included only if they have the ﬁve years of historical sales data necessary to compute SG.g.. Frankel. Lee / Journal of Accounting and Economics 25 (1998) 283–319 for years t!1. the eﬀect is again not monotonic.. Bernard. More importantly.g. R>G Consistent with prior studies — e. reﬂecting the compounding eﬀects of pevious year forecast errors. Analyst overoptimism (pessiR>G mism) relative to future reported earnings results in more positive (negative) FErr values. Fairﬁeld.. FErr is computed by subtracting each ﬁrm’s R>G predicted year t#i ROE (FROE ) from the actual reported ROE in period R>G t#i (ROE ). 1995. this R> relation is not monotonic across the quintiles.and threeyearahead biases are somewhat higher. not on forecasted or actual ROEs. forecast errors are deﬁned with the opposite sign (e. The two. as well as our use of the longterm growth rate to estimate threeyearahead FROEs. Ret36 is the average threeyear buyandhold return for each portfolio. 1988. Similarly. In most prior studies. overlyoptimistic: FErr is R>G positive for all quintiles in all three Panels.306 R. Speciﬁcally. but is otherwise ﬂat.. 1988). and does not hold in one. O’Brien. However. and FErr is the average forecast error in the year R>G t#i forecasts. Panel A shows a negative relation between B/P and FErr — analysts are most overoptimistic in low B/P ﬁrms. Fried and Givoly (1982).M. Statistical signiﬁcance of the diﬀerence in means between extreme quintiles is determined using Monte Carlo techniques.g. Consistent with prior studies (e. Analyst overoptimism increases sharply for the highest SG quintile.
Finally. Based on the last row in Table 6. Lee / Journal of Accounting and Economics 25 (1998) 283–319 307 through Q4. but drop sharply for the top SG quintile. Model 3 shows that when » is much higher (lower) than » . Similarly. Because the independent variables are all expressed in terms of percentile ranks.5% lower than those of the bottom percentile B/P ﬁrms.0%. Model 1 shows that low (high) B/Ps are generally associated with ex post analyst optimism (pessimism). To compute its percentile rank. but the relation is nonlinear. Frankel. Since the typical ﬁrm earns an ROE of 13%. Table 6 reports the result of 15 annual crosssectional regressions of realized analyst forecast errors on each of these four ﬁrm characteristics (year t"1977—1991).3%. who also ﬁnd that ﬁrm rankings on past growth measures (both earnings and sales) do not result in a strong systematic return diﬀerentials in intermediate portfolios. each variable is sorted as of the end of June in year t and assigned to percentiles. . these diﬀerences appear substantial and economically signiﬁcant. Table 6 shows that the relation between these four variables and the subsequent analyst forecast error is robust over time. To facilitate interpretation of these results and to reduce the eﬀect of outliers. The independent variables are R> SG. Newey—West (1987) tstatistics based on timeseries variations in the annual estimates (bottom row) indicate statistical signiﬁcant at the 1% level for all four models. This result is consistent with Dechow and Sloan (1997). Model 2 shows that high (low) past sales growth (SG) is positively associated with excessive optimism (pessimism). the intermediate portfolio returns to ¸tg rankings are monotonic. Panel C shows that the OP portfolios are also related to analyst forecast errors. C. the independent variables are expressed in terms of their percentile ranks. or ¸tg. OP.R. B/P. and unlike SG. The dependent variable for each regression is FErr . Panel D shows that ¸tg has strong predictive power for returns. the top B/P ranked percentile ﬁrms have forecasted ROE errors that are 2. we can compare their estimated coeﬃcients. analysts tend to be too optimistic (pessimistic). The diﬀerence between top and bottom percentile OP ﬁrms is 7. Similarly. analyst forecasts tend to be more overoptimistic for high OP ﬁrms in all forecast horizons. Overall. Model 4 shows that high ¸tg ﬁrms tend to have overly optimistic forecasts. the diﬀerence in forecasted ROE errors for the top and bottom SG percentiles is 4.C. while the diﬀerence between top and bottom percentile Ltg ﬁrms is 7. Panel D shows that analyst forecasts tend to be more overoptimistic for high ¸tg ﬁrms.3%. The sign of the estimated coeﬃcient is correct in 55 out of 60 individual cases. To evaluate the robustness of these relations over time. Taken together. As predicted by the mispricing hypothesis.M. Panels A through D show that all four variables appear to have some predictive power for crosssectional diﬀerences in longterm analyst forecast errors. the evidence suggests SG is related to analyst forecast errors in the direction predicted by the mispricing hypothesis.
09*** 0.081 5.046 2. Numbers in the All Years row represent timeseries means and Newey—West (1987) tstatistics based on timeseries variation in the annual estimates.023 !1. 206 173 261 387 631 684 694 684 729 691 669 678 678 756 779 All Years 15 years .081 4.086 4. tstat Coeﬀ.20** 0.029 !1.56* 0.01*** 0.37*** 0.150 9.067 3. tstat Coeﬀ.09*** !0.115 7.87*** 0. B/P is the booktomarket ratio.060 3.47*** 0.Table 6 The relation between analyst forecast errors and ex ante ﬁrm characteristics This table reports the results of 15 (t"1976—1990) crosssectional regressions of realized analyst forecast errors on four ﬁrm characteristics.27 0.56*** 0.064 3.028 !0.11*** 0.000 !0.036 2.077 4.19*** 0.89*** 0.27*** 0.104 !0.008 !0.57*** Obs. Forecast errors (the dependent variable) are computed by subtracting actual returnsonequity (ROEs) in period t#2 from predicted t#2 ROEs obtained from I/B/E/S consensus earning forecasts available prior to June 30 of year t.83 !0.049 3.24*** 0.45 !0.029 2. 5% and 10% levels respectively.049 3. tstat Mean tstat 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 RK(B/P) !0.088 5.047 2.090 5.016 !0. Speciﬁcally. SG is the ﬁveyear percentage growth in sales from period t!6 to t!1. and * signify statistical signiﬁcance at the 1%.92*** 0.002 0.001 0. where » is an EBO value derived using current I/B/E/S analyst consensus forecasts and » is a similar EBO value measure derived using historical ROEs.19*** 0.073 6.169 11.01*** Model 4 RK(¸tg) 0. tstat Coeﬀ.19*** 0. Model 1 Year t Coeﬀ.25*** 0.009 !0.54* 0.40*** 0.081 5.029 1.040 2.84** !0. tstat Coeﬀ.02*** 0.113 6.068 4. RK(.140 8. tstat Coeﬀ.128 7.09*** 0. tstat Coeﬀ.016 0.69*** Model 2 RK(SG) 0.96*** 0.007 !0. **.71** !0.043 7.93*** 0. The independent variables are computed as follows.10 0.043 2.66*** 0.92 0.) is a percentile rank operator.64 0.033 !1.67 0.003 0. To compute percentile ranks.68 !0.078 4.447 0. where B is the reported book value per share for the ﬁscal year ended in year t!1 and P is the stock price at the end of June in year t.13*** 0.070 5. tstat Coeﬀ. tstat Coeﬀ.074 !4. ***.022 1.25 !0.059 3.025 !2.00*** !0. tstat Coeﬀ.010 !0.12*** 0.015 1.82 !0.050 !3.48*** 0.100 6.069 4.97*** 0.085 5.647 !0.40*** 0.048 2.143 7. tstat Coeﬀ.03*** 0.001 0.02*** 0.00*** Model 3 RK(OP) 0.005 0. tstat Coeﬀ. OP is a measure of analyst optimism derived using Edward—Bell—Ohlson (EBO) fundamental value measures.96*** 0.035 1.061 4.017 0.036 2.020 !1.013 !0.130 !8. the independent variables are sorted as of the end of June each year and assigned to percentiles.50*** 0.20 0. ¸tg is the consensus longterm earnings growth forecast from I/B/E/S as of May of year t. tstat Coeﬀ.28*** !0.016 0. OP"(» !» )/"» ".71*** 0.37*** 0.68*** 0. tstat Coeﬀ.59* !0.011 0.20 0. tstat Coeﬀ.77 0.
tstat Coeﬀ. and assigned to percentiles. tstat Coeﬀ.002 0. ** and * signify onetailed statistical signiﬁcance at the 1%.52 0.62 1.80 0.065 3.062 3. the top and bottom 1% forecasted error each year are omitted.117 0.018 1. 3.087 0.00 0.009 0.49 0.57 !0.91 0. tstat Coeﬀ.03 0.150 0. OP"(» !» )/"» ".056 2.072 3.012 !0.023 1.01 0. tstat Coeﬀ.46 0.132 7.71 0.048 2.042 1.17 0. tstat Coeﬀ.031 1. tstat Coeﬀ. Lee / Journal of Accounting and Economics 25 (1998) 283–319 Table 7 Predicting analyst forecast errors using multiple ﬁrm characteristics 309 This table reports the results of 15 (t"1976—1990) multiple crosssectional regressions of realized analyst forecast errors on four ﬁrm characteristics. Year t Coeﬀ.33 0.74 0. the independent variables are sorted as of the end of June each year.46 0.044 2.35 0.94*** RK(B/P) RK(OP) !0. To reduce the eﬀect of outliers. ¸tg is the consensus longterm earnings growth forecast from I/B/E/S.86 0.056 2.017 0.051 3. Numbers in the All Years row represent timeseries means and Newey—West (1987) tstatistics based on timeseries variation in the annual estimates.26 0. tstat Coeﬀ.88 0.002 !0.25 0.C. where B is the reported book value per share for the ﬁscal year ended in year t!1 and P is the stock price at the end of June in year t. 5% and 10% levels. tstat Coeﬀ.62 !0.47 0.5*** 22.03 0.06 0.013 !0.43 !0.029 1.097 0.031 1.023 0.050 0.042 2.032 1.89 0.55*** RK(¸tg) 0.66*** 11.006 0. tstat Coeﬀ.010 1. To compute percentile ranks.073 !4.065 3.076 4.075 0.055 0.014 0. tstat Coeﬀ.R.5*** 31.22 0.13*** 41.58 0.042 2.183 9. SG is ﬁve year percentage growth in sales from period t!6 to t!1.056 3.62 0.020 1.4*** 13.023 1.31** 1.075 0.00 0.92 !0.010 0.015 0.51 0.64 9.020 !1.067 3.061 0.019 !0. where » is an EBO value derived using current I/B/E/S analysts consensus forecasts and » is a similar EBO value measure derived using historical ROEs.62 0.87 0.064 3.46 0.10 0.049 0.52 0.35 !0.7*** 8.44 0.M. RK(.012 0.16 0.035 5.041 2. tstat coeﬀ.24 0.66*** 3.59 0.0*** 206 173 261 387 631 684 694 684 729 691 669 678 678 756 779 All Years 0.62 0.030 1.055 2.64 0.010 0.068 3.033 1.2*** 18.195 0. The independent variables are computed as follows.040 2.) is a percentile rank operator. Forecast errors for each ﬁrm (the dependent variable) are computed by subtracting actual returnsonequity (ROEs) in period t#2 from predicted t#2 ROEs obtained from I/B/E/S consensus earning forecasts.089 3. C.66 !0.82 0.019 0.98 0.042 2.36 0.13 !0.92 0.23 0.025 1.109 5.020 !0.020 0.030 !1.010 !0.096 5.5*** 13.56 0. ***.9*** 16.43 0. tstat Coeﬀ.054 Fstatistic Obs. tstat Mean tstat 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 RK(SG) 0.007 0.70 0.26*** 15 years .014 0. B/P is the booktomarket ratio. respectively.56 R 0. Frankel.73 0.029 2.119 5.82 0.48 0.079 4.36 0.52 !0. tstat Coeﬀ.096 5.37 0.45 0.000 0. tstat Coeﬀ.09 0.074 4.
SG and ¸tg provide greater explanatory power for forecast errors than B/P.5. we use scaled decile rankings for .and threeyearahead buyandhold returns for our sample ﬁrmyears on scaled decile ranks of ME. and ¸tg variables to compute a predicted forecast error. portion of the error in the consensus I/B/E/S forecast is predictable each year. From these annual regressions. but consistent. K . Speciﬁcally.6. GR\ (4) The parameters L . Large positive (negative) values of PErr correspond to excessively optimism (pessimism) forecasts. we derive estimated coeﬃcient weights for each variable. we regressed the one.310 R. the consistency of the coeﬃcients suggests a potential trading strategy. the predicted forecast error for ﬁrm i in portfolio formation year t is: PE " L # K RK(SG )# K RK(BP )# K RK(OP ) PPGR GR\ GR\ GR\ # K RK(¸tg ). we regress forecast errors realized in year t!1 on percentile ranks of SG.5% in 1982. RK(. and PErr.M. OP. the NeweyWest (1987) timeseries tstatistics in the last row show that after controlling for the other variables. We then apply these estimated coeﬃcients to each ﬁrm’s period t!1 B/P. With each variable conditioned on the others. Table 7 reports the results of annual multiple regressions of these forecast errors on ranked percentiles of B/P. 5. Although the predictable portion is not large. K . SG.4%. The R for the annual regressions ranges from a low of 1% in 1979 to a high of 19. SG. and the annual Fstatistics indicate signiﬁcance in 12 out of 15 years. Lee / Journal of Accounting and Economics 25 (1998) 283–319 5. Speciﬁcally. a modest. Proﬁting from the forecast error To exploit the predictable component of the I/B/E/S forecast error. Predictability of analyst forecast errors We now assess the amount of crosssectional variation in forecast errors that is explained by combining our four variables. Since three past years of data are necessary to estimate PErr. Frankel. B/P. In fact. Table 8 presents returns to a PErr strategy and compares the results to returns from other investment strategies. and ¸tg. we ﬁrst estimate annual crosssectional regressions of the form presented in Table 7. To construct this table. K and K are estimated from rolling crosssectional regressions based on year t!4 information and actual year t!1 earnings. The average R is 7. Following Bernard and Thomas (1990) and Dechow and Sloan (1997). B/P oﬀers little incremental contribution to the model.C. We examine this possibility in the next section. » /P. B/P. OP. Clearly. the sample period for this test is 1979—1992. OP and ¸tg from year t!4. C.) is the percentile rank operator. OP.
Clearly. The results are the same as those reported in Fig. and 27. Over the next 36 months. over threeyear holding periods.7% over the next 36months.and threeyearahead returns.5%. a PErr strategy. the PErr strategy performs about as well as » /P and B/P. Recall from Table 7 that during these years our prediction model .C. over a 36month holding period. For the combined strategy. cumulative returns are the average returns from selling ﬁrms in the top quintile (high PErr ﬁrms) and buying ﬁrms in the bottom quintile (low PErr ﬁrms). The results show that none of the strategies perform particularly well over oneyear holding periods. Over the next 12 months. Lee / Journal of Accounting and Economics 25 (1998) 283–319 311 all independent variables. This evidence extends the results in Table 4 by illustrating that » /P has incremental predictive power controlling for both ME and B/P. Returns to the B/P and » /P strategies are based on buying ﬁrms in the top quintile and selling ﬁrms in the bottom quintile each year. Model 6 in both panels shows that the PErr strategy enhances the predictive power of » /P. 1. 2 presents a comparison of the cumulative monthly returns produced by four alternative trading strategies: a B/P strategy. Frankel.R. PErr has signiﬁcant predictive power for one. In the early years (pre1982). indicating that the PErrbased strategy has incremental explanatory power to » /P. Both results are statistically signiﬁcant at the 1% level. These results show that » /P has signiﬁcant incremental power to predict crosssectional returns in both one. For each calendar year. Fig. a » /P strategy. Table 8 shows that a decilebased PErr strategy yields approximately 4% over the next 12months. and a combined strategy. this combined strategy yields a cumulative return of 45. we buy (sell) ﬁrms that are simultaneously in the top (bottom) » /P quintile and the bottom (top) PErr quintile. even controlling for ME and B/P. 2 shows that the highest returns are produced by the combined strategy. For the PErr strategy. the independent variables are assigned in descending order to deciles. Fig. PErr. but provides a better picture of the robustness of each strategy over time. Indeed. but underperforms the » /P strategy. the PErr strategy outperforms the B/P strategy. and then scaled so that they range from zero (for the lowest decile) to one (for the highest decile). C. Table 9 reports the yearbyyear results of implementing each strategy. Models 5 and 6 in Table 8 evaluate the incremental contribution » /P and PErr controlling for B/P and ME. However. » /P is the most important variable in explaining crosssectional returns. the » /P strategy does better than the combined strategy.and threeyearahead regressions. Model 5 in Panel B shows that over 36 months. In addition. » /P and the combined strategy all outperform the B/P strategy. This test has less statistical power to detect abnormal returns.M. This approach allows the regression coeﬃcients to be interpreted as estimates of the return to a zero investment portfolio with a long position in the stocks in the highest decile and a short position in the stocks of the lowest decile.
040*** 0.196*** 0.37 0. Panel A: Oneyearahead returns Model 1 2 3 4 5 6 Intercept 0.155*** 0. The parameters L .039*** !0.and threeyearahead buyandhold returns. Lee / Journal of Accounting and Economics 25 (1998) 283–319 Table 8 Relative importance of ex ante ﬁrm characteristics in returns prediction This table presents estimated coeﬃcients from regressions of one. and an optimism measure (OP"(» !» )/"» ").539*** !0. All independent variables are assigned in descending order to deciles.) is a percentile rank operator.029* !0.688*** — — 5 0. RK(.168*** — 2 0.15 0.051 0.00 1. **.365*** — — 4 0. since 1982. the PErr strategy appears to add noise without adding additional predictive power. where » is similar to » .312 R. and then scaled so that they range from zero (for the lowest decile) to one (for the highest decile). ***. booktoprice (BP). Market value of equity (ME) and stock price (P) are as of June 30 of each year t.09 0. The sample consists of 11.013 6 0.029 !0. For ﬁrm i and period t.468*** 0.147*** 0.352*** — 0.12 0.C. » is a fundamental value estimate derived using current I/B/E/S consensus forecasts available prior to June 30 of year t. we require each ﬁrm’s ﬁveyear past sales growth (SG). Book value (B) is book equity per share in calendar year t!1. respectively.861 ﬁrmyears between 1976 and 1992 for ﬁrms with all data available. including dividends and any liquidating distributions.035** Panel B: Threeyearahead returns 1 0. * Denote signiﬁcance at the 1%.47 performed poorly. Yet the strategy is not . However.241*** 0. 5% and 10% levels.343*** !0.538*** — 0. Thus.030** !0.031** — 0.and threeyearahead stock returns on various ex ante ﬁrm characteristics.053* — — — — 0.176*** BP ME » /P PErr Adj. the combined strategy outperformed » /P every year.01 0. and the predicted analyst forecast error (PErr).151*** 0.023 — — — — 0.026 3 0. and K are estimated from rolling crosssectional regressions based on year t!4 information and year t!1 reported earnings.051*** — — !0. K K K . the predicted forecast error is PErr " L # K RK(SG )# K RK(BP )# K RK(OP )# K RK(¸tg ).341*** 0. R (%) 0. The 36month return to the combined strategy is consistently positive through up and down markets. The return cumulation period begins at the end of June in each year t.186*** 0. Frankel. valuetoprice (» /P). markettobook ratio (BP).277*** 0. longterm consensus earnings growth forecast (¸tg). Large positive (negative) values of PErr correspond to predictions of excessive overoptimism (pessimism). PErr is the predicted error in the year t consensus forecast of year t#2 ROEs.83 1. The independent variables are market value of equity (ME). Dependent variables are the one. using a twotailed ttest. but derived using historical earnings rather than analyst forecasts.83 2.019 — — — — 0.03 1.07 0.19 0.370*** — — !0. GR GR\ GR\ GR\ GR\ This predicted error is estimated using information available prior to June of year t. Speciﬁcally.013 0.042*** — — !0. C.M.
and maintaining these investments until the end of the indicated month. C. In 1981. For the combined PErr and » P strategy. B is book equity per share in calendar year t!1.R. 6. particularly over . without risk. This ﬁgure shows the cumulative returns from several alternative trading strategies. portfolios are formed at the end of June by sorting ﬁrms into quintiles on the basis of B/P. P is price per share at the end of June 30 of year t. ﬁrms are included in the long (short) portfolio if they are simultaneously in the top » /P quintile and the bottom PErr quintile. Lee / Journal of Accounting and Economics 25 (1998) 283–319 313 Fig. the combined strategy lost 66%. The PErrbased strategy is simialr. due to large losses in a few stocks. PErr is the predicted error in consensus analyst forecasts for year t#2 returnonequity (ROE). this graph depicts the cumulative buyandhold returns produced by buying ﬁrms in the top quintile and selling ﬁrms in the bottom quinitle at the beginning of July. and » /P. The sample period is 1979—1991 (year t"1970 to 1991). A comparison of cumulative (buyandhold) returns from alternative trading strategies. For the B/P and » /P based strategies. Summary In this study. Frankel. » is a fundamental value estimate based on the consensus analyst forecast.C. Our results show that » /P is a reliable predictor of crosssectional returns. we operationalized an analystbased residual income model and used the resulting valuetoprice (» /P) ratio to examine issues related to market eﬃciency and the predictability of crosssectional stock returns. PErr. particularly over a oneyear horizon.M. estimated as of June 30 of year t. except ﬁrms in the top quintile are sold and ﬁrms in the bottom quintile are purchased. 2. Each year.
670 0.136 0.589 0.130 0.102 0.155 !0.228 0.529 0.074 0.066 — !0.314 R.250 !0.and » /Pbased strategies (Panels A and C.038 0. where a trading position maybe either a long position or a short position.176 0.32** 164 to 392 !0.038 — 0.159 0.675 0.397 0.133 !0.457 5.075 !0. P is price of the stock at the end of June 30 in year t. Each year.273 0.06*** 164 to 393 . ***. Number of ﬁrms indicates the highest and lowest number of trading positions taken per year. C.099 0.363 0.027 0. » is the EBO fundamental value measure based on the consensus analyst forecast.077 !0.239 0.083 0.071 0.083 !0. Numbers in the Mean row represent timeseries means of the annual returns.046 1. PErr.002 !0.062 0. Lee / Journal of Accounting and Economics 25 (1998) 283–319 Table 9 Yearbyyear returns to various trading strategies This table reports the cumulative oneyear and threeyear buyandhold returns from alternative trading strategies. For B/P.184 — 0.110 0.322 0.247 !0.019 0.129 0.092 !0. and maintaining these investments for either 12 or 36 months.909 1.190 !0.057 0.M.171 0.008 0.C.23 164 to 442 0.109 0. The PErrbased strategy (Panel B) is similar.499 0.204 0.431 !0.325 0.55*** 164 to 393 !0.322 0.402 — — 0.052 0.127 0.118 0.745 0.091 0.124 0. B is book equity per share in calendar year t!1.659 0.221 !0.141 0. respectively).113 0.264 0.188 0.104 0.022 0.043 0. ** and * signify onetailed statistical signiﬁcance at the 1%. and » /P.228 3.295 0.158 0.349 4.62 164 to 442 0.244 0.286 0. portfolios are formed at the end of June by sorting ﬁrms into quintiles on the basis of B/P.25 0.456 0.090 0.214 !0. 5% and 10% levels respectively.02 164 to 442 0.199 !0.328 0.038 1.052 0. the table values represent the average cumulative equalweighted returns produced by buying ﬁrms in the top quintile and selling ﬁrms in the bottom quintile at the beginning of July of each year.147 — 0.017 0. with Newey—West (1987) correction for serial correlation.07 48 to 149 3year 0.044 0.553 0.274 1.393 0.05 0.103 0.072 0. except ﬁrms in the top quintile are sold and ﬁrms in the bottom quintile are bought.447 !0. Panel A B/P (high—low) Year t 78 79 80 81 82 83 84 85 86 87 88 89 90 91 Mean tstat Number of ﬁrms 1year 3year Panel B PErr (low—high) 1year 3year Panel C » /P (high—low) 1year 3year Panel D Combined PErr and » /P 1year !0.105 0.244 0.294 0.199 0.410 0. Reported tstatistics are based on timeseries variations in the annual means.016 0. For the combined PErr and » /P strategy (Panel D).263 3.189 0.062 0.234 0.349 0.063 0. PErr is the predicted error in consensus analyst forecasts of year t#2 returnonequity.005 !0.261 0.411 0.183 0.40*** 48 to 149 !0. Frankel.027 — !0.338 !0.091 — 0.176 !0.343 0.171 0. ﬁrms are included in the long (short) portfolio if they are simultaneously in the top » /P quintile and the bottom PErr quintile.
Our evidence suggests that ﬁrm value estimates based on a residual income model may be a useful starting point for predicting crosssectional stock returns. Herzberg (1998) shows a strong partial price correction in the ﬁrst month after the strategy is implementable (see Exhibit 12). We ﬁnd that crosssectional errors in threeyearahead consensus forecast are predictable. Dechow et al. but it does not explain why returns remain high in years two and three. Over the next 12months. In addition. has received signiﬁcant attention. the predictive power of » /P is comparable to that of B/P. Our ﬁnding that prices converge to value estimates gradually over longer horizons (beyond 12months) is puzzling. we ﬁnd some evidence that analysts tend to be more overlyoptimistic in ﬁrms with higher past sales growth and P/B ratios. future work may focus on alternative mechanical models of earnings prediction (e. or beta. to the conservative nature of our tests. 1997).g. Because of its importance in estimating » . in part. However. Our implementation of the residual income model is simple. we implement this strategy with a 5 to 6 week lag — we use IBES forecasts publicly available by the third week of May and form portfolios as of June 30th. and show this estimate has predictive power for crosssectional returns. Moreover. ﬁrm size.. The bias in longterm analyst .M. Lee / Journal of Accounting and Economics 25 (1998) 283–319 315 longer horizons. This ability to predict longterm returns is not attributable to B/P. we develop an estimate of the prediction error in longterm forecasts (PErr). One explanation is that the price convergence to value is a much slower process than prior evidence suggests. Our ﬁrstyear results therefore do not capture the shortterm proﬁts from the ﬁrst 6 weeks. about the process by which information about longterm fundamentals is impounded in price. C. Our results suggest that superior return prediction may result from adopting a more complete valuation approach. This explanation suggests our ﬁrst year returns should be higher. Much recent research has focused accountingbased ratios that exhibit predictive power for stock returns. » /P has much stronger predictive power than B/P.R. The eﬀect may be due. However. Future studies may also lead to reﬁnements in other key parameters of the model. and in particular. in particular.C. Frankel. over the next 36months. and leaves much room for improvement. Speciﬁcally. This possibility raises interesting questions about the eﬃciency of the market. While we focus on an analystbased valuation model. we show this predictive power is incremental to a » /P strategy.. In a recent replication of our results. The B/P ratio. we ﬁnd stronger evidence of overoptimism in ﬁrms with higher forecasted earnings growth (¸tg) and higher forecasted ROEs relative to current ROEs (OP). Combining these variables in a prediction model. including forecasted dividend payout ratios and crosssectional variations in discount rates. We hope that our ﬁndings will encourage further research along these lines. and a combined strategy yields the highest returns. we also investigate the reliability of longterm I/B/E/S consensus earnings forecasts.
In addition. Ken French. it is diﬃcult to understand why arbitrage forces do not eliminate this pricing anomaly more quickly. FROE and FROE ]. Jay Shanken (the referee). We gratefully acknowledge the ﬁnancial support of the QGroup and the KPMG Peat Marwick Foundation (Lee). unlike returns to a B/P strategy. C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 forecasts is not realized until two or three years into the future.316 R. Larry Brown. Richard Sloan. the University of Minnesota.M. 7. We derive these R> R> . Steve Merritt. Alternatively. Pat O’Brien.P. Earnings forecasts used in this paper are provided by I/B/E/S. Using I/B/E/S forecasts to derive future ROE estimates Our implementation of the Edwards—Bell—Ohlson (EBO) formula requires three future ROE forecasts [FROE . We also thank Jeﬀ Abarbanell.C. Tom Dyckman. Our evidence on the predictability of longterm forecast errors in consensus forecasts is consistent with this longterm mispricing hypothesis. Bhaskaran Swaminathan. James Myers provided expert research assistance. Ohio State University. we acknowledge that high » /P ﬁrms may still be riskier than low » /P ﬁrms in some other. Kothari (Editor). the University of Rochester. returns to a » /P strategy exhibits a pattern of lower shortterm returns and higher longterm returns. We ﬁnd that high » /P ﬁrms generally have lower market betas. For Further Reading The following references are also of interest to the reader: Dechow et al. (1998) and Fama and French (1955). as yet unidentiﬁed. » /P may be yet another proxy for crosssectional risk diﬀerences. This pattern is diﬃcult to reconcile with a risk explanation. and Yale University for helpful suggestions. whose many insights helped to bring the residual income valuation model to life. Bruce Lehman. Appendix A. Our tests control for two obvious sources of potential risk: the B/P ratio and ﬁrm size. dimension. and analysts appear to revise their longterm forecasts only gradually over time. S. Kent Daniel. Despite these concerns. the University of Oregon. Frankel. a Michigan MBA student. John Core. deserves credit for ﬁrst identifying an apparent market anomaly with this trading rule during a class exercise. an anonymous referee. Harvard University. We leave this question to future research. so sensitivity to overall market movements is an unlikely explanation for their higher subsequent returns. Jim Bodurtha. However. Dartmouth College. Georgetown University. and workshop participants at Cornell University. Acknowledgements We thank the late Victor Bernard.
R R\ R Step 2: Estimating FROE and B . Year t refers to the year of portfolio formation. Journal of Accounting and Economics 14. J. Since yearend book values are dependent on current year ROEs. R> R R R> R R> and B .. We also require that all sample ﬁrms R> R> have a twoyearahead consensus forecast [F½2]. Forecasted ROE for year t is then computed as the year t consensus forecast.R.. Frankel. 1181—1207. Bernard.. Abarbanell. We then compute FROE R> and B analogously: R> FROE "F½2/[(B #B !1)/2].S. book value reduces the chance of an extremely low denominator. Abarbanell.. B "B [1#FROE (1!k)]. 1995. S.D. 147—165.C. C. J. we have: FROE "F½1/[(B #B )/2]. Journal of Financial Economics 43. Step 1: Estimating FROE and B . 1997. J. 341—372. J. We require that all sample ﬁrms have R a oneyearahead I/B/E/S consensus EPS forecast [F½1]. Notationally. Barber. Kothari. R> R> References Abarbanell.. R> R> R B "B [1#FROE (1!k)].. R. University of Michigan.. . V. Where a longterm earnings growth Step 3: Estimating FROE R> R> and B as follows: estimate [¸tg] is available. Journal of Financial Economics 38. Bernard. we use FROE to proxy for FROE . 1995.M. R R\ R\ B "B [1#FROE (1!k)]. Ball. V.. Lyon. J. Tests of analysts’ overreaction/ underreaction to earnings information as an explanation for anomalous stock price behavior. we use a sequential process to estimate future ROEs.. we compute FROE R> R> FROE "[F½2(1#¸tg)]/[(B #B )/2]. 1992..P. The steps in the process are listed below. R> R> R> Where ¸tg is not available. Lee / Journal of Accounting and Economics 25 (1998) 283–319 317 future ROEs from I/B/E/S consensus EPS estimates. We then use FROE and the dividend payout ratio (k) to derive the ending book value for R year t. 1991. rather than yearend. divided by the average book value per share during year t!1. 79—107. Is the U. Do analysts’ earnings forecasts incorporate information in prior stock price changes?. January. Detecting longrun abnormal stock returns: The empirical power and speciﬁcation of test statistics. Journal of Finance 47. Shanken. stock market myopic? Working paper. Problems in measuring portfolio performance: an application of contrarian investment strategies. Use of the average. B.M.
. 1995. Berkeley. Contrarian investment. 1991.. extrapolation...F. Journal of Finance. 1994. University of Michigan and Cornell University.D. P. E. forthcoming. Financial Analysts Journal July/Aug. Journal of Investing. D.. S. Griﬃn. Edwards.. January.. and implications for ﬁnancial statement analysis.F. J. 49—67.. Journal of Finance 47. C. determinants of booktomarket ratios. Security analyst superiority relative to univariate timeseries models in forecasting quarterly earnings. 1998.L. A.... Predicting analysts’ earnings surprise. The crosssection of realized stock returns: the preCompustat evidence. Implementing EBO/EVA analysis in stock selection. Keon. Quinn.C.. V. Sloan. French.B. Journal of Finance 44. Journal of Accounting and Economics 9. Investment performance of common stocks in relation to their price earnings ratios: a test of the eﬃcient market hypothesis. University of California Press..L. Fundamentals and stock returns in Japan.G. 85—107. Hagerman. 1998... Business cycle variation in earnings forecasts and common stock returns. Zmijewski... Lee / Journal of Accounting and Economics 25 (1998) 283–319 Basu. Westerﬁeld.. 427—465. 1997. 3—27. Working paper.... Herzberg. Schwager. and stock returns. Accountingbased valuation methods.. January.. International Journal of Forecasting 9. P.. S. 1961. 1579—1593. Industry costs of equity. Jr. K. 61—87. A. Fairﬁeld. 1977..A. 1541—1578.. J.M. Givoly. Hutton.P. French.. Fama. L. P.. Accounting diversity and international valuation. S. L. 153—193.. D. Chan.C. Han. Journal of Accounting and Economics 4. 23—31. G. P/E.. J. Valuation and clean surplus accounting for operating and ﬁnancial activities. S. Another look at the crosssection of expected returns. R. Lakonishok. J. E. L.. 689—731. Working paper. 1994. Dechow.M. Y. 1739—1764..F. 663—682. Size and booktomarket factors in earnings and returns... and risk. Fama. K..318 R.. 1987b. R.J. Kothari. J... CA. 1995.D. University of Michigan. .R. E. 1989. 1993.K. W. 1998. Hamao. Financial analysts’ forecasts of earnings: a better surrogate for market expectations. Shanken. 1994.. Bell. Lakonishok.. Sloan.M..E. Forthcoming. Working paper. K. E. Brown. Journal of Financial Economics 43. P. Betting on trends: intuitive forecasts of ﬁnancial risk and return. 1997. French. Kothari. Frankel. Journal of Financial Economics 43. Contemporary Accounting Research 11. D. Davis. R. W. Earnings yields. market values. Journal of Finance 46. Bernard.H. 185—224. 1987a.. J.. An empirical assessment of the residual income valuation model.. Ohlson. R. Feltham. 1992.Y. Working paper. 301—340.W.P. 26. Journal of Accounting Research 25. The Journal of Investing 45—53. September. P/B and the present value of future dividends. Dechow.F. Warner. 1982. forthcoming.M. Brown. Returns to contrarian investment: tests of the naive expectations hypothesis.. 1994.M. L. Journal of Finance 32. J....R. Brown.D. Measuring longhorizon security price performance. Jaﬀe. Shleifer. Vishny. 355—371.C. Keim. G.C. R. Richardson. Journal of Finance 50. 1994.. Journal of Finance 48. Lee.R. 1995..A. The crosssection of expected stock returns. C. An information interpretation of ﬁnancial analyst superiority in forecasting earnings.. J. Journal of Accounting Economics. K.. M. 135—148.D. V. Journal of Finance 49. Journal of Financial Economics 43. DeBondt.. The Theory and Measurement of Business Income. Frankel. E. Fama.A. University of Chicago and University of Nebraska at Omaha. Daniel. P. 1997. Fried. R. Mande..B. Sloan.L. 1995. R. M.
Contemporary Accounting Research 7. and Dividends in Security Valuation. R. J. 347—372. May. 1938. Journal of Financial Economics 7. 1996. J. T. Lehman. 1984. Kahneman. Ramaswamy. C. 1124—1131.C. Lee / Journal of Accounting and Economics 25 (1998) 283–319 319 LaPorta. Journal of Finance 52.. 1979. Penman. 1995. 1991. The eﬀect of personal taxes and dividends on capital asset prices. Summary ﬁnancial statement measures and analysts’ forecasts of earnings. 859—874. Judgment under uncertainty: heuristics and biases. O’Brien. Earnings. 1988. Journal of Accounting and Economics 15. K. Annual survey of economic theory: the theory of depreciation. Improved methods for tests of longrun abnormal stock returns.. Econometrica 6. Measuring wealth CA Magazine April. J.... Working paper. 53—83.A.L.H. Myers. October..A. 1995. A synthesis of security valuation theory and the role of dividends. Ohlson. R... J.C. 648—676. U. 1997. R. Penman. Litzenberger. 1990. Lee. 1993. Good news for value stocks: further evidence on market eﬃciency. Contemporary Accounting Research 6. 32—37. Journal of Finance 51. G. 1998. Peasnell. Journal of Accounting and Economics 10.C.. C.L. B. University of California at Berkeley.M. P.A. Sougiannas.. Ou. Lakonishok. Ohlson.M.. dividend policy. 1996. 661—687. Penman.. B. A comparison of dividend. forthcoming. 263—282. cash ﬂow. Shleifer. and earnings approaches to equity valuation. .. The theory of value and earnings. 1715—1742. S. B.. Lyon. Lee. 1998. Tversky. 1998.H. J. forthcoming. A. Journal of Finance 53. D. What is the intrinsic value of the Dow? Journal of Finance. Working paper. Preinreich.H.R. C.C. Tsai.M. 1994. Some formal connections between economic values and yields and accounting numbers. Vishny. Financial statement analysis and the evaluation of booktomarket ratios. February. Review of Quantitative Finance and Accounting 3. 1982..D. C. Book Values. Frankel. cash ﬂows.A. and present value relations: building blocks of dividend policy invariant cash ﬂows... Working paper. S. Earnings. K. Berkeley. J.. T... Barber. A. Analysts’ forecasts as earnings expectations. Stober. 1992.C. Santa Clara University and the University of California at Berkeley. J. and earnings. and University of Illinois at UrbanaChampaign. Expectations and the crosssection of stock returns..H.. 219—241. 163—195... and an introduction to the BallBrown analysis. Ohlson. LaPorta. Swaminathan. 1—19.. S...M.. R. Journal of Business Finance and Accounting 361—381. Science 185. A synthesis of equity valuation techniques and terminal value calculations for the dividend discount model. Contemporary Accounting Research 11.
This action might not be possible to undo. Are you sure you want to continue?