The Center for Research in Security Prices

Working Paper No. 456

February 1999

University of Chicago
Graduate School of Business

“Forecasting Profitability and Earnings"

Eugene F. Fama
Graduate School of Business, University of Chicago

Kenneth R. French
Sloan School of Management, MIT

This paper can be downloaded without charge from the

Social Science Research Network Electronic Paper Collection:
http://papers.ssrn.com/paper.taf?abstract_id=40660

Forthcoming: The Journal of Business Vol. 73, No. 2.

An index to the Working Papers in the Center for Research in Security Prices Working Paper Series is located
at: http//gsbwww.uchicago.edu/fac/finance/papers/
 First draft: June 1997
Revised: February 1999
Not for Quotation: Comments Welcome

FORECASTING PROFITABILITY AND EARNINGS

Eugene F. Fama and Kenneth R. French*

Abstract

There is a strong presumption in economics that, in a competitive environment, profitability is mean

reverting. We provide corroborating evidence. In a simple partial adjustment model, the estimated rate of mean

reversion is about 40 percent per year. But a simple partial adjustment model with a uniform rate of mean

reversion misses rich non-linear patterns in the behavior of profitability. Specifically, we find that mean reversion

is faster when profitability is below its mean and when it is further from its mean in either direction. We also

show that the mean reversion in profitability produces predictable variation in earnings.

*
Graduate School of Business, University of Chicago (Fama); Sloan School of Management, MIT, and NBER
(French). The comments of Douglas Diamond, Elizabeth Gordon, Douglas Hanna, and a referee are gratefully
acknowledged.
 There is a strong presumption in economics that profitability is mean reverting. For example, Stigler

(1963, p.54) states,

“There is no more important proposition in economic theory than that, under competition, the rate of
return on investment tends toward equality in all industries. Entrepreneurs will seek to leave relatively
unprofitable industries and enter relatively profitable industries…”

These standard economic arguments imply that in a competitive environment, profitability is mean

reverting within as well as across industries. Other firms eventually mimic innovative products and technologies

that produce above normal profitability for a firm. And the prospect of failure or takeover gives firms with low

profitability incentives to allocate assets to more productive uses.

Mean reversion in profitability implies that changes in profitability and earnings are to some extent

predictable. There is a large literature, mostly in accounting, that attempts to identify predictable variation in

earnings and (less commonly) profitability. The evidence is, however, difficult to judge, for three reasons.

(i) Some early studies that produce tantalizing suggestions of predictability provide no formal tests [for

example, Beaver (1970), Lookabill (1976), Brooks and Buckmaster (1976)].

(ii) When formal tests are provided, they are often based on time-series models fit separately to

individual firms. To enhance power, the tests are restricted to firms with long earnings histories; requiring 20

years of data is common. Although 20 years is a long period in a firm’s life, 20 observations on annual earnings

produce imprecise estimates of a time-series model. As a result, evidence that suggests economically interesting

predictability is typically statistically weak [for example, Lev (1969), Freeman, Ohlson and Penman (1982)].

(iii) Cross-section regressions of changes in profitability or earnings on lagged changes and other

variables seem to produce more reliable evidence of predictability [Freeman, Ohlson, and Penman (1982), Ou

and Penman (1989), Collins and Kothari (1989), Easton and Zmijewski (1989), Elgers and Lo (1994), Basu

(1997)]. With the exception of Elgers and Lo (1994), however, the standard errors of the regression slopes in

these tests are not adjusted for the correlation of regression residuals across firms. In effect, the standard errors

are based on the patently unrealistic assumption that there is no correlation across firms in current changes in

profitability and earnings (due, for example, to current macro-economic or industry shocks) beyond that absorbed

by lagged predictor variables.

 2

We take a different tack that preserves the power of the cross-section tests but produces inferences that

allow for residual cross-correlation. In the spirit of Fama and MacBeth (1973), we forecast profitability and

earnings with year-by-year cross-section regressions, and we use the average slopes and their time-series standard

errors to draw inferences. This approach allows us to use large samples, an average of 2304 firms per year. And

the year-by-year variation in the slopes, which determines the standard errors of the average slopes, includes the

effects of estimation error due to the correlation of the residuals across firms.

Allowing for residual cross-correlation is important. The average slopes from our year-by-year cross-

section regressions are essentially equivalent to the slopes from pooled time-series cross-section regressions that

include annual dummies that allow the average values of the variables to change through time. Pooled time-series

cross-section regressions are a popular choice for solving the power problem in times-series tests on individual

firms [for example, Freeman, Ohlson, and Penman (1982), Fairfield, Sweeney, and Yohn (1996)]. And the

inference problem created by residual cross correlation is typically ignored. Without showing the details, the

Fama-MacBeth standard errors of our average regression slopes (which account for residual cross correlation)

are typically two to five times larger than the OLS standard errors from pooled time-series cross-section

regressions.

Accurate inferences are important, but our main contribution is substantive. Confirming standard

economic arguments, we find strong evidence that profitability (measured as the ratio of year t earnings before

interest to total book assets, Yt/At) is mean reverting. In a simple partial adjustment model, the rate of mean

reversion is about 40 percent per year. But a simple partial adjustment model with a uniform rate of mean

reversion does not do justice to the rich patterns in the behavior of profitability. Specifically, we find that the rate

of mean reversion is higher when profitability is far from its mean, in either direction. The rate of mean reversion

is also higher when profitability is below its mean.

Like changes in profitability, changes in earnings are predictable. When we restrict the explanatory

variables to lagged earnings changes, we confirm the early qualitative evidence of Brooks and Buckmaster (1976)

that changes in earnings tend to reverse from one year to the next, and large changes of either sign reverse faster
 3

than small changes. We also confirm the formal evidence of Elgers and Lo (1994) that negative changes in

earnings reverse faster than positive changes.

For the mostpart, the existing literature does not examine the links between the predictability of

profitability and the predictability of earnings. This is true even in the rare papers that examine both variables

[for example, Beaver (1970), Ball and Watts (1972)]. In contrast, like Freeman, Ohlson, and Penman (1982)

and Lev (1983), our hypothesis is that much of what is predictable about earnings is due to the mean reversion

of profitability. And our tests confirm the hypothesis. Our results thus imply that real-world forecasts of

earnings (for example, by security analysts) should incorporate the mean reversion in profitability. There is,

however, predictable variation in earnings beyond that captured by our model for profitability. Specifically,

negative changes in earnings and extreme changes seem to reverse faster than predicted by our specific non-linear

model for profitability.

We begin (section I) with estimates of a simple partial adjustment model for profitability in which the

rate of mean reversion is a constant, that is, it does not depend on how profitability deviates from its expected

value. Section II fine-tunes the model to allow for non-linear mean reversion. Section III tests for predictable

changes in earnings. Section IV concludes.

I. A First-Pass Partial Adjustment Model for Profitability

As a first test for predictable variation in profitability, we estimate, for each year t from 1964 to 1995,

a simple cross-section partial adjustment regression for the change in profitability from t to t+1,

(1a) Yt+1/At+1 – Yt/At = a + b[Yt/At – E(Yt/At)] + c[Yt/At – Yt-1/At-1] + et+1,

(1a) CPt+1 = a + bDFEt + cCPt + et+1.

At is a firm’s total book assets at the end of year t; Yt is earnings before interest and extraordinary items but after

taxes; Yt/At is our measure of profitability; E(Yt/At) is its expected value; CPt = Yt/At – Yt-1/At-1 is the change

in profitability from t-1 to t; and DFEt = Yt/At – E(Yt/At) is the deviation of profitability from its expected value.
 4

To simplify the notation, we omit the firm subscript that should appear on the regression variables and residuals,

and the year subscript that should appear on the regression coefficients.

We use a two-step approach to estimate (1). Each year t, we regress Yt/At for the firms in our sample

on variables meant to capture differences across firms in expected profitability. We then use the fitted values

from this first-stage regression as the proxy for E(Yt/At) in the cross-section estimate of (1) for year t.

We use three variables to explain expected profitability in the first-stage regressions. (i) An old

hypothesis is that dividends have information about expected earnings because firms target dividends to the

permanent component of earnings [Miller and Modigliani (1961)]. Thus one of our proxies for expected

profitability is Dt/BEt, the ratio of year t dividends to the book value of common equity at the end of the year.

(ii) Fama and French (1999) find that firms that do not pay dividends tend to be much less profitable than

dividend payers. To capture any resulting non-linearity in the relation between dividends and expected

profitability, our second variable is a dummy, DDt, that is 0.0 for dividend payers and 1.0 for non-payers. (iii)

Since the market value of a firm, Vt, is the current value of all future net-cash-flows, we use the market-to-book

ratio, Vt/At, to pick up variation in expected profitability missed by the dividend variables. [Additional proxies

for expected profitability tried but dropped for lack of explanatory power include the log of total assets (a

measure of size) and the ratio of depreciation expense to total assets (a measure of capital intensity).] The proxy

for expected profitability, E(Yt/At), in (1) is then the fitted value from the cross-section regression,

(2) Yt/At = d0 + d1Vt/At + d2DDt + d3Dt/BEt + εt.

The tests exclude financial firms and utilities. Financial firms and utilities are highly regulated during

much of our sample period, and regulation may produce unusual behavior of profitability. We are also concerned

that influential observations might dominate the regressions. The variables in (1) and (2) are scaled by assets

or book equity. This can create influential observations when At and BEt are close to zero. To address this

problem, firms with total assets less than $10 million and book equity less than $5 million are not used in the

tests. These exclusions still leave us with hefty cross-sections, an average of 2304 firms per year.
 5

We use the average slopes and the time-series standard errors of the average slopes in (1) and (2) to draw

inferences. A benefit of this approach is that the standard errors of the average slopes include estimation error

due to the correlation of the regression residuals across firms. Autocorrelation in the slopes from the year-by-year

regressions is also a problem. The higher-order autocorrelations are random about zero, but first-order

autocorrelations are sometimes large, around 0.5. We could adjust the standard errors of the average slopes for

the estimated autocorrelation of the annual slopes. But with just 32 observations on the slopes for 1964-95,

autocorrelation estimates are imprecise; the standard errors are around 0.18. We use a less formal approach to

allow for the autocorrelation of the annual regression slopes. With first-order autocorrelations around 0.5, the

variances of the average slopes, calculated assuming serial independence of the annual slopes, are too small by

about 50%, and the standard errors of the average slopes should be inflated by 40%. Thus we require t-statistics

around 2.8, rather than the usual 2.0, to infer reliability.1

Part A of Table 1 shows average slopes from the first-stage regressions of Yt/At on Vt/At, DDt, and

Dt/BEt. All three variables have information about expected profitability, E(Yt/At). The positive average slope

on Dt/BEt is 8.44 standard errors from zero. The negative average slope on DDt confirms that the relation

between profitability and dividends is non-linear; the expected profitability of firms that do not pay dividends

is 0.025 lower (t = - 8.27) than predicted by the relation between Yt/At and Dt/BEt. The strong positive average

slope on Vt/At (t = 10.82) is consistent with the hypothesis that the market-to-book ratio captures variation in

expected profitability missed by the dividend variables.

Part B of Table 1 shows average slopes from second-stage estimates of the partial-adjustment model (1)

that do not constrain the slopes on expected profitability, E(Yt/At), and observed profitability, Yt/At. The partial-

adjustment model predicts that the Yt/At slope is negative, the E(Yt/At) slope is positive, and the two slopes are

equal in absolute value. The average Yt/At and E(Yt/At) slopes, -0.40 and 0.37, confirm these predictions. These

1
Elgers and Lo (1994) use annual cross-section regressions to model the autocorrelation of changes in earnings. Somewhat
uniquely, they also use the time-series standard errors of average regression slopes to allow for the correlation of residuals
across firms. But they do not adjust the standard errors of the average slopes for the autocorrelation of the annual slopes.
And they examine a much narrower range of models for the predictability of earnings.
 6

average slopes are also quite precise (t = -18.00 and 13.55). All this is striking evidence that profitability is mean

reverting.

The lagged change in profitability, CPt = Yt/At – Yt-1/At-1, is included as an explanatory variable in (1)

to test whether the mean reversion captured by the partial adjustment term, DFEt = Yt/At – E(Yt/At), is the sole

source of predictable variation in profitability. Table 1 shows that when the lagged change in profitability, CPt,

is used alone to explain CPt+1, the slope on CPt is strongly negative; on average, the change in profitability from

t to t+1 reverses 30 percent (t = -10.84) of the lagged change. But allowing for mean reversion, that is, including

Yt/At and E(Yt/At) in the regression, moves the CPt slope close to zero, -0.09. Still, the average CPt slope is -4.72

standard errors from zero. Thus there is small but statistically reliable negative autocorrelation in changes in

profitability beyond what can be explained by the partial adjustment term.

It is worth noting that the strong evidence that the change in profitability is predictable from the lagged

change is testimony to the power of the cross-section regressions. In a typical time-series test using 20 years of

annual changes, the standard error of the autocorrelation of successive changes is (1/19)½ = 0.23. Thus, the

coefficient -0.30, which is -10.84 standard errors from zero in the cross-section regressions, would be only 1.30

standard errors from zero in a time-series regression with 20 observations. This is the generic power problem

in tests that attempt to identify predictable variation in profitability (or earnings) using time-series models fit to

individual firms.

Our estimate of the rate of mean reversion of profitability, in the neighborhood of 0.40 (40 percent per

year) is similar to the median of the estimates for individual firms in Lev (1969). He fits separate partial

adjustment models to 20 years of time-series data for individual firms. The resulting estimates of the rate of mean

reversion are imprecise; the median is only 1.71 standard errors from zero. The cross-section partial adjustment

regressions in Fairfield, Sweeney, and Yohn (1996) also produce rates of mean reversion near 0.4 (0.35). But

they test the reliability of the estimates with out-of-sample forecasts that do not adjust for the correlation of the

forecast errors across firms. We corroborate these earlier estimates, but with more precise methods.
 7

Finally, simple economic stories say that competitive forces push profitability toward a common

economy-wide mean. In the partial adjustment model (1), however, we allow E(Yt/At) to vary across firms.

Cross-sectional differences in expected profitability can occur for several reasons. First, even with perfect

competition, differences in risk produce differences in expected profitability. Second, our measure of

profitability, Yt/At, is a noisy proxy for true economic profitability. For example, even if all firms share the same

true expected profitability, systematic differences between the historical and replacement costs of assets create

systematic differences in E(Yt/At). Third, differences in expected profitability can be the result of monopoly

rents.

How are the estimates of (1) affected if we assume all firms revert toward one overall equilibrium level

of expected profitability? With this assumption, the partial adjustment model in (1a) simplifies to

(3) Yt+1/At+1 – Yt/At = a + bYt/At + c[Yt/At – Yt-1/At-1] + et+1,

and profitability reverts to the grand mean at the rate -b. Table 1 shows that the estimated rate of mean reversion

from (3) is 31 percent per year (t = 14.53). The somewhat higher rates of mean reversion produced by (1) then

suggest that (2) captures meaningful differences across firms in expected profitability.

II. A Non-Linear Partial-Adjustment Model for Profitability

Though they attempt no formal inferences, Brooks and Buckmaster (1976) present evidence that changes

in earnings are likely to reverse from one year to the next, the reversals are stronger for extreme changes of either

sign, and they are stronger for negative changes. Elgers and Lo (1994) formally confirm the last result. These

papers focus on changes in earnings. Given our hypothesis that the predictability of earnings should be largely

due to mean reversion in profitability, it is interesting to test whether there is similar non-linearity in the behavior

of profitability. To this end, we expand the partial adjustment model (1) as,

(4a) CPt+1 = a + (b1 + b2NDFEDt + b3NDFEDt*DFEt + b4PDFEDt*DFEt)DFEt

+ (c1 + c2NCPDt + c3NCPDt*CPt + c4PCPDt*CPt)CPt + et+1,

(4b) = a + b1DFEt + b2NDFEt + b3SNDFEt + b4SPDFEt

 8

+ c1CPt + c2NCPt + c3SNCPt + c4SPCPt + et+1.

NDFEDt, PDFEDt, NCPDt, and PCPDt are dummy variables. NDFEDt is 1.0 when DFEt (the deviation of

profitability from its expected value) is negative, and zero otherwise; PDFEDt is 1.0 when DFEt is positive;

NCPDt is 1.0 when CPt (the change in profitability from t-1 to t) is negative; PCPDt is 1.0 when CPt is positive.

The derived variables in (4b) are negative deviations of profitability from its expected value (NDFEt), squared

negative deviations (SNDFEt), squared positive deviations (SPDFEt), negative changes in profitability (NCPt),

squared negative changes (SNCPt), and squared positive changes (SPCPt). In brief, b2, b3, and b4 measure non-

linearity in the mean reversion of profitability, that is, in the speed of adjustment of profitability to its expected

value. And c2, c3, and c4 measure non-linearity in the autocorrelation of changes in profitability.

Table 1 suggests that there is non-linearity in the autocorrelation of changes in profitability similar to

that observed by Brooks and Buckmaster (1976) and Elgers and Lo (1994) for changes in earnings. When we

estimate (4b) without the mean reversion variables (that is, suppressing b1 to b4), we find that changes in

profitability tend to reverse (the CPt slope is negative), negative changes tend to reverse faster than positive

changes (the NCPt slope is negative), and reversal is stronger for more extreme changes (the SNCPt slope is

positive and the SPCPt slope is negative). But most of these results are statistically weak. Only the SNCPt slope

breaks our 2.8 standard error reliability barrier.

We are more interested in whether the autocorrelation of changes in profitability can be attributed to

mean reversion in the level of profitability. Table 1 shows that when we add the mean reversion variables, that

is, we estimate the full version of (4b), the slopes c1 to c4 on the autocorrelation variables move toward zero, and

c1, c2, and c4 are less than 1.25 standard errors from zero. The slope c3 on SNCPt falls by about two-thirds in

the full regression, but it is 3.45 standard errors from zero.

In the end, non-linear mean reversion is a fairly complete story for the predictable variation in

profitability. Allowing for non-linearity pushes the slopes on Yt/At and E(Yt/At) (the two components of the

linear partial adjustment term, DFEt) toward zero, but they remain opposite in sign, strikingly close in absolute

value (-0.15 and 0.14), and more than five standard errors from zero. And the rate of mean reversion is highly
 9

non-linear. Mean reversion is stronger when profitability is below its mean; the slope on NDFEt is -0.12 (t =

-2.71). Mean reversion is also stronger when profitability is further from its mean; the slopes on the quadratic

terms SNDFEt and SPDFEt, 1.02 and -1.89, are more than 4.5 standard errors from zero.

From an economic perspective, it is plausible that the rate of mean reversion is higher when profitability

is below its mean and when it is far from its mean in either direction. When profitability is low, the prospect of

failure or takeover gives firms an incentive to allocate assets to more productive uses. And the incentive is

stronger the further profitability is below its mean. Conversely, the incentives of other firms to mimic the

products and technologies of their rivals increase when the rivals are more profitable. It is also possible, however,

that non-linear mean reversion, and mean reversion itself, trace in part to accounting decisions. For example,

Basu (1997) argues that bias toward conservative reporting leads firms to report losses quickly but to spread

gains over longer periods. Such a tendency could help explain why profitability reverts more quickly when it is

low.

III. Predicting Earnings

The existing predictability literature focuses primarily on earnings rather than profitability. Moreover,

the literature is largely agnostic about the economic forces that cause earnings to be predictable. Notable

exceptions are Freeman, Ohlson, and Penman (1982) and Lev (1983). They argue that competitive forces

produce mean reversion in profitability, which is then the source of predictable variation in earnings. This is also

the stance adopted here. We now examine whether changes in earnings are predictable and how much of the

predictability traces to the non-linear mean reversion of profitability.

Table 2 shows estimates of (4) in which the dependent variable is the (scaled) change in earnings, CEt+1

= (Yt+1 – Yt)/At, rather than the change in profitability, CPt+1 = Yt+1/At+1 – Yt/At. The new regression is,

(5a) CEt+1 = a + (b1 + b2NDFEDt + b3NDFEDt*DFEt + b4PDFEDt*DFEt)DFEt

+ (c1 + c2NCEDt + c3NCEDt*CPt + c4PCEDt*CPt)CEt + et+1

(5b) = a + b1DFEt + b2NDFEt + b3SNDFEt + b4SPDFEt

 10

+ c1CEt + c2NCEt + c3SNCEt + c4SPCEt + et+1.

The explanatory variables in the first line of (5a) and (5b) are those used in (4a) and (4b) to capture the mean

reversion of profitability. The new variables are in the second line of (5a) and (5b). They are designed to pick

up autocorrelation in earnings changes left unexplained by the mean reversion of profitability. NCEDt and PCEDt

in the second line of (5a) are dummy variables. NCEDt is 1.0 when CEt (the change in earnings from t-1 to t) is

negative and zero otherwise. PCEDt is 1.0 when CEt is positive. The derived variables in the second line of (5b)

are negative changes in earnings (NCEt), squared negative changes (SNCEt), and squared positive changes

(SPCEt). Thus, c2, c3, and c4 are meant to pick up non-linearity in the autocorrelation of changes in earnings.

Table 2 says that there is indeed non-linearity in the autocorrelation of earnings changes. When we use

only the lagged change, CEt, to predict CEt+1, the slope is -0.14. This corroborates time-series evidence [Beaver

(1970), Ball and Watts (1972)], that the linear autocorrelation of successive changes in earnings is weak; earnings

seem to behave much like a random walk. In our cross-section tests, however, the autocorrelation of -0.14 is -

4.33 standard errors from zero. Thus, even in a linear model, the deviation of earnings from a random walk is

reliable. When we add negative changes in earnings to the linear model, the slope on CEt turns slightly positive,

but the NCEt slope is strongly negative, -0.48 (t = -8.81). This is in line with the evidence in Elgers and Lo

(1994) that the reversal tendency of changes in earnings is more reliable for negative changes. But our tests also

say that reversal is stronger for more extreme changes in earnings of either sign. When we add squared positive

and squared negative changes in earnings (SPCEt and SNCEt) to the regressions that include CEt and NCEt, the

slope on NCEt remains strongly negative (-0.42, t = -7.64) and the SNCEt and SPCEt slopes, 1.71 and -0.63, are

4.33 and -5.44 standard errors from zero.

In short, there is a rich non-linear pattern in the autocorrelation of successive changes in earnings. The

reversal tendency of earnings changes is stronger for more extreme changes of either sign. And reversal is

stronger when earnings have declined. All of this is formal corroboration of the early informal evidence on

predictable variation in earnings in Brooks and Buckmaster (1976).

 11

We are interested in whether the autocorrelation of earnings changes can be attributed in whole or in part

to non-linear mean reversion in the level of profitability. Table 2 shows that when we add the linear mean

reversion variables Yt/At and E(Yt/At) to the regressions that also include the four autocorrelation variables (CEt,

NCEt, SNCEt and SPCEt), there is strong evidence that mean reversion leads to predictable variation in

profitability; the slopes on Yt/At and E(Yt/At), -0.32 and 0.46, are -13.07 and 13.19 standard errors from zero.

But the average slopes on the autocorrelation variables remain strong (more than four standard errors from zero).

Allowing for the non-linear mean reversion behavior of profitability (adding DFEt, SNDFEt, and SPDFEt to the

regressions) moves the slopes on most of the autocorrelation variables closer to zero, but the CEt, NCEt, and

SPCEt slopes are still more than four standard errors from zero. Thus, it seems that the non-linear behavior of

profitability is far from a complete story for the predictable variation in earnings.

Since the specific functional form we use to model profitability is surely not a complete story for the

mean reversion of profitability, it is not surprising that it misses some of the predictability of earnings changes.

But there is another reason to question the specification of the earnings change regressions. Unconditional

expected earnings growth, E(Yt+1 – Yt)/At = E(CEt+1), surely differs across firms. As a result, some of the

apparent forecast power of the explanatory variables in the CEt+1 regressions is probably driven by differences

in unconditional expected earnings changes rather than by true predictability of earnings. In contrast, since

profitability cannot drift forever up or down, the unconditional expected value of the change, CPt+1 = Yt+1/At+1

– Yt/At, is probably close to zero for most firms. Thus, the profitability regression (4) is probably better specified

than the earnings change regression (5).

There is evidence that variation across firms in unconditional expected earnings growth is a problem in

the estimates of (5). Though noisy, the lagged change in earnings is a likely candidate to pick up variation across

firms in E(CEt+1). Since reversal seems to be the general characteristic of earnings changes, the positive slope

on CEt in the full version of (5), 0.18 (t = 5.56), suggests that this variable does identify variation in

unconditional expected earnings growth.

 12

One might get cleaner estimates of the predictability of changes in earnings by modeling variation across

firms in expected earnings growth. But our view is that profitability, not earnings, is the interesting economic

variable. And the interesting predictability issue is whether competitive forces produce mean reversion in

profitability.

IV. Conclusions

Standard economic arguments say that in a competitive environment, profitability is mean reverting. Our

evidence is in line with this prediction. In a simple partial adjustment model, the rate of mean reversion is about

40 percent per year. But the mean reversion of profitability is highly non-linear. Mean reversion is faster when

profitability is below its mean and when it is far from its mean in either direction.

There is also predictable variation in earnings. Much of it traces to the mean reversion of profitability.

An important practical implication of this result is that forecasts of earnings (for example, by security analysts)

should exploit the mean reversion in profitability. There does, however, seem to be predictable variation in

earnings beyond that captured by our model for profitability. In particular, negative changes in earnings and

extreme changes seem to reverse faster than predicted by our specific model for the mean reversion of

profitability.
 13

References

Ball, Ray, and Ross Watts, 1972, Some time-series properties of accounting income, Journal of Finance 27,
663-682.

Basu, Sudipta, 1997, The conservatism principle and the asymmetric timeliness of earnings, Journal of
Accounting and Economics 24, 3-37.

Beaver, William H., 1970, The time series behavior of earnings, Journal of Accounting Research 8
(Supplement), 62-99.

Brooks, LeRoy D., and Dale A. Buckmaster, 1976, Further evidence of the time series properties of accounting
income, Journal of Finance 31, 1359-1373.

Collins, Daniel W., and S.P. Kothari, 1989, An analysis of intertemporal and cross-sectional determinants of
earnings response coefficients, Journal of Accounting and Economics 11, 143-181.

Easton, Peter D., and Mark Zmijewski, 1989, Cross-sectional variation in the stock market response to
accounting earnings announcements, Journal of Accounting and Economics 11, 117-141.

Elgers, Peter T., and May H. Lo, 1994, Reductions in analysts’ annual earnings forecast errors using information
in prior earnings and security returns, Journal of Accounting Research 32, 290-303.

Fairfield, Patricia M., Richard J. Sweeney, and Teri Lombardi Yohn, 1996, Accounting classification and the
predictive content of earnings, Accounting Review 71, 337-355.

Fama, Eugene F., and Kenneth R. French, 1999, Dividends, debt, investment, and earnings, manuscript, Graduate
School of Business, University, of Chicago, January.

Fama, Eugene F., and James D. MacBeth, 1973, Risk, return, and equilibrium: Empirical tests, Journal of
Political Economy 81, 607-636.

Freeman, Robert N., James A. Ohlson, and Stephen H. Penman, 1982, Book rate-of-return and prediction of
earnings changes: An empirical investigation, Journal of Accounting Research 20, 639-653.

Lev, Baruch, 1969, Industry averages as targets for financial ratios, Journal of Accounting Research 7
(Autumn), 290-299.

Lev, Baruch, 1983, Some economic determinants of time-series properties of earnings, Journal of Accounting
and Economics 5, 31-48.

Lookabill, Larry L., 1976, Some additional evidence on the time series properties of accounting earnings, The
Accounting Review 51 (October), 724-738.

Miller, Merton H., and Franco Modigliani, 1961, Dividend policy, growth, and the valuation of shares, Journal
of Business 34, 411-433.

Ou, Jane A., and Stephen H. Penman, 1989, Accounting measurement, price-earnings ratio, and the information
content of security prices, Journal of Accounting Research 27, 111-152.
 14

Stigler, George J., 1963, Capital and Rates of Return in Manufacturing Industries, Princeton University Press,
Princeton, NJ.
 Table 1 -- Regressions to Explain the Level of and Change in Profitability: 1964-95, 32 Years
The regressions are run for each year t, from 1964 to 1995, using NYSE, AMEX, and NASDAQ firms on
Compustat with data for the year on all variables in any regression and with total assets greater than $10 million and book
common equity greater than $5 million. The table shows means (across years) of regression intercepts (Int) and slopes, and
t-statistics for the means t(Mn), defined as the mean divided by its standard error [times-series standard deviation of the
coefficient divided by (32)½]. The table also shows averages (across years) of the means and standard deviations (Std) of
the regression variables.
The variables (and Compustat data items) are as follows. At is total book assets (6). Dt is dividends paid during
fiscal year t (21). DDt is a dummy that is 1.0 when dividends are zero and zero otherwise. Yt is earnings before
extraordinary items (18), plus interest expense (15), plus (when available) income statement deferred taxes (50) and
investment tax credit (51). Book common equity, BEt, is At minus liabilities (181) minus preferred stock [taken to be, in
order and as available, redemption value (56), liquidating value (10), or par value (130)], plus balance sheet deferred taxes
and investment tax credit (35). Vt, the total value of the firm, is its common stock price (199) times shares outstanding at
the end of fiscal year t (25), plus At minus BEt. E(Yt/At), the proxy for expected profitability, is the fitted value from the
profitability regression for year t, summarized in Part A. DFEt is Yt/At – E(Yt/At). NDFEt is DFEt when DFEt is negative
and zero otherwise. SNDFEt is the square of DFEt when DFEt is negative and zero otherwise. SPDFEt is the square of DFEt
when DFEt is positive and zero otherwise. CPt is Yt/At – Yt-1/At-1. NCPt is CPt when CPt is negative and zero otherwise.
SNCPt is the square of CPt when CPt is negative and zero otherwise. SPCPt is the square of CPt when CPt is positive and
zero otherwise.

Part A: Regressions to Explain the Level of Profitability, Yt/At

Means and t-Statistics for the Means of the Year-by-Year Regression Coefficients
Int Vt/At DDt Dt/BEt R2
Mean .047 .025 -.025 .20 .25
t(Mn) 22.694 10.816 -8.270 8.44 8.36
Means and Standard Deviations of the Regression Variables
Yt/At Vt/At DDt Dt/BEt
Mean .076 1.444 .329 .031
t(Mn) .074 .953 .440 .059

Part B: Regressions to Explain the Change in Profitability, CPt+1 = Yt+1/At+1 – Yt/At

Means and t-Statistics for the Means of the Year-by-Year Regression Coefficients
Int Yt/At E(Yt/At) NDFEt SNDFEt SPDFEt CPt NCPt SNCPt SPCPt R2
Mean -.003 -.40 .37 -.09 .21
t(Mn) -1.129 -18.00 13.55 -4.72 8.54
Mean .017 -.31 -.13 .19
t(Mn) 8.192 -14.53 -7.29 7.74
Mean -.006 -.30 .10
t(Mn) -3.934 -10.84 5.05
Mean -.011 -.08 -.11 1.58 -.22 .15
t(Mn) -5.134 -2.27 -1.58 5.44 -1.22 6.11
Mean -.006 -.15 .14 -.12 1.02 -1.89 -.04 .03 .55 -.20 .24
t(Mn) -2.248 -5.43 5.19 -2.71 4.52 -5.80 -1.21 .62 3.45 -1.13 9.57
Means and Standard Deviations of the Regression Variables
CPt+1 Yt/At E(Yt/At) NDFEt SNDFEt SPDFEt CPt NCPt SNCPt SPCPt
Mean -.005 .076 .076 -.019 .004 .002 -.004 -.019 .004 .003
Std .072 .074 .029 .049 .037 .030 .068 .046 .056 .044
 Table 2 -- Regressions to Explain the Change in Earnings: 1964-95, 32 Years
The regressions are run for each year t, from 1964 to 1995, using NYSE, AMEX, and NASDAQ firms on
Compustat with data for the year on all the variables in any regression. Part A of the table shows means (across years) of
the regression intercepts (Int) and slopes, and t-statistics for the means t(Mn), defined as the mean divided by its standard
error [the times-series standard deviation of the regression coefficient divided by (32)½]. Part B shows averages (across
years) of the means and standard deviations (Std) of the regression variables. E(Yt/At), the proxy for expected profitability,
is the fitted value from the profitability regression for year t, summarized in Part A of table 1. DFEt is Yt/At - E(Yt/At).
NDFEt is DFEt when DFEt is negative and zero otherwise. SNDFEt is the square of DFEt when DFEt is negative and zero
otherwise. SPDFEt is the square of DFEt when DFEt is positive and zero otherwise. CEt is (Yt - Yt-1)/At-1. NCEt is CEt
when CEt is negative and zero otherwise. SNCEt is the square of CEt when CEt is negative and zero otherwise. SPCEt is
the square of CEt when CEt is positive and zero otherwise.

Part A: Average Slopes for Regressions to Explain the Change in Earnings, dYt+1/At = (Yt+1-Yt)/At

Int Yt/At E(Yt/At) NDFEt SNDFEt SPDFEt CEt NCEt SNCEt SPCEt R2
Mean .011 -.14 .05
t(Mn) 6.886 -4.33 3.80

Mean .002 .04 -.48 .09

t(Mn) 1.318 1.16 -8.81 5.92

Mean .000 .19 -.42 1.71 -.63 .12

t(Mn) .065 5.83 -7.64 4.33 -5.44 6.00

Mean -.008 -.32 .46 .24 -.23 1.49 -.67 .18

t(Mn) -3.247 -13.07 13.19 7.63 -5.12 4.20 -5.84 7.44

Mean -.015 .05 .17 -.30 1.14 -2.17 .18 -.25 .50 -.48 .20
t(Mn) -5.940 1.59 3.31 -4.91 4.00 -6.25 5.56 -5.94 1.53 -4.80 7.80

Part B: Means and Standard Deviations of the Regression Variables

CET+1 Yt/At E(Yt/At) NDFEt SNDFEt SPDFEt CEt NCEt SNCEt SPCEt

Mean .010 .076 .076 -.019 .004 .002 .012 -.015 .003 .004
Std .073 .074 .029 .049 .037 .030 .072 .041 .040 .053