American Finance Association

Wiley
Some Time Series Properties of Accounting Income
Author(s): Ray Ball and Ross Watts
Source: The Journal of Finance, Vol. 27, No. 3 (Jun., 1972), pp. 663-681
Published by: Wiley for the American Finance Association
Stable URL: http://www.jstor.org/stable/2978268
Accessed: 10-10-2015 08:01 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact [email protected].

American Finance Association and Wiley are collaborating with JSTOR to digitize, preserve and extend access to The Journal
of Finance.

http://www.jstor.org

This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

SOME TIME SERIES PROPERTIES OF ACCOUNTING INCOME

RAYBALLANDRoss WATTS*
I. INTRODUCTION
THE TIMESERIESbehavior of corporate incomes has generated considerable
interest. The early work of Little [28] suggests that successive changes in
corporate incomes in the U.K. are independent. Subsequent empirical work
suggests that the same conclusion applies to corporate incomes in the U.S.
This article applies different methods from those of previous researchers,but
the conclusion is unchanged. We conclude that, in general, measured annual
accounting incomes for U.S. corporations follow either a submartingale or
some very similar process.1 Because of small sample problems this conclusion
is based on mean and median results; no attempt has been made to determine
whether specific'firms are systematic outliers.
The conclusion that corporate incomes are submartingales has important
implications for forecasters and researchers in accounting and finance. For
example, it implies that attempts to smooth corporate incomes in the manner
suggested in the accounting and finance literatures are not successful. It also
affects the interpretationof the growth and decline of firms.
OF INCOME
II. THE SMOOTHING

The accounting and finance literatures are replete with the suggestion that
accountants smooth the incomes of firms. That is, it is commonly hypothesized
that accountants manipulate their income-measuring techniques in order to
soften the effect of hard times upon income and, conversely, in order to
diminish the extent to which good times are contemporaneously reflected in
income.

The hypothesis that accountants do smooth income and the belief that they
should smooth income have existed for decades. Hepworth [25], Gordon [21,
223, Gordon et al. [23], Schiff [36], Dopuch and Drake [14], Copeland
[10], Copeland and Licastro [11], and Gagnon [20] all investigate income
* Graduate School of Business, University of Chicago and Graduate School of Management,
University of Rochester,respectively.The authors are indebted to Philip Brown, Eugene Fama,
Nicholas Gonedes, Merton Miller, Robert Officerand a referee for their assistance.An earlier
version of this paper was deliveredat the Workshopin AccountingResearchand the Workshop
in Finance at the Universityof Chicagoon April 23, 1968.
1. Let Y1,Y2 . . . be random variables with expectations.Then the sequence {YJ is a submartingaleif
for all t,
E(Yt l I Yol . .. I Yt) >. Yt
where E is an expectationoperator.
The martingaleis a specific case of a submartingale.
The sequence{YJ} is a martingaleif
E(Yt+l I Yo .*., Yt) = Yt

for ail t.

663
This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

664

The Journal of Finance

smoothing. To our knowledge, no study has addressed the possible futility of

alleged smoothing practices.
As it is presented in the literature, smoothing is an attempt to reduce the
variance of income around its expectation. Income is assumed to be generated
by a process whose expectation is constant or is a deterministic function of
time.2 For example, Schiff writes [36, p. 66]:
Someyearsago, Bouldingreferredto the "homeostasis
of the balancesheet-that there
is some desiredquantityof all the variousitems in the balancesheet, and that any
disturbanceof this structureimmediatelysets in motion forces which restore the
status quo." It can be suggestedthat we now have a "homeostasisof earningsper
share"and that the applicationof generallyacceptedaccountingprinciplesfacilitates
the reportingof earningsper sharein a constantor risingpattern.
Gordon writes [22, p. 223]:
If the variationof the observationsaroundthe curveare smaller[whenaccountants
adopta specificaccountingpractice],incomesmoothinghas beenthe consequence....
Copeland says essentially the same thing [10, p. 102]:
Incomesmoothinginvolvesthe repetitiveselectionof accountingmeasurementor
reportingrulesin a particularpattern,the effectof whichis to reporta streamof incomewith a smallervariationfromtrendthan wouldotherwisehave appeared.
These and other sources imply that the expectation of income is a function
of time or is constant. Smoothing implies a return to good times, on average,
after bad times, during which income decreases are artificially reduced by
smoothing practices. It implies that many increases in income are also
temporary, and can therefore be smoothed in order to avoid the impression
of permanence.
Apart from ignoring the substantial evidence that the market can decide
for itself whether in fact income changes are permanent or temporary,3 the
smoothing of income by the accounting profession would seem to ignore the
possibility that good times are not followed, on average, by bad times. A submartingale implies that a firm is stuck with good and bad times (deviations
of realized incomes from expectations) when they occur since a submartingale,
by definition, is a process in which any one observation becomes the basis
for the expectation of the next. Given the uncertainty of the world, there is
always a probability distribution around the expectation. But the behavior
of the expectation of a submartingale over time would be such as to make
nonsense of the notion of income smoothing.4
are randomvariableswith expectations,the assumptionis
for all t,
E(Yt+l I Yo...* Yt) = V
where V is a constant or
E(Yt+ IYo...,Yt) =f(t+ 1)
2. If Y1,Y2

. ..

where f(t + 1) is some function of t + 1, and which incorporates trends in expectations in an

analogous fashion to a submartingale.
3. See the theory and evidence summarized by Fama [16]. Some specific studies of market
ingenuity in reacting to information, such as income, are [2] and [18].

4. Attemptingto smooth an underlyingsubmartingalemechanismwould increasethe varianceof

future income changes,simply by making the probabilitydistributionof changesasymmetric.

This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

Propertiesof AccountingIncome

665

One could initially be tempted to argue that this is not true, because the
evidence of independence of income changes applies only to measured income
numbers-which are observed after accountants have decided upon which
measurementrules to use. This argumentdoes not go far enough, for smoothing
of a series generated by a process with an expectation that is constant or a
deterministic function of time produces another series of the same form.5
Since the observed series is not of this form, smoothing of this type does not
appear to occur. If accountants try to smooth in the manner which we outline,
then they attempt a futile exercise."
III. GROWTHAND DECLINE
The interpretation of growth and decline (and their extreme counterparts,
survival and failure) depends heavily upon the income-generating process.
Growthin a martingalemechanism occurs as frequently as decline; and either,
once experienced, is permanent, on average. Growth in a process whose expectation is constant over time is nonexistent.7
The implications of income variability for the survival of a firm depend
upon the process which generates income. A constant-expectation finite-variance process implies the relative insignificance of variability in income.
Deviations of income from the expectation are then once-and-for-all increments or decrements in the value of the firm. Value changes are relatively
small-the value of the firm changes in the order of the size of the deviation
in income, given the (known) expectation of the process. If the expectation
of income is known and stationary (and it surely would become known), then
investors face very little risk. It is difficult to see why individual firms would
ever fail; even the possibility of gambler's ruin and its consequential reorganization costs seem to be avoidable by borrowing when the expectation
of income is constant. While borrowing is not without cost, it is then difficult
to see the ex ante importance of variability to the valuation of the firm, conditional upon a (known) long-run expectation.
In contrast, a process whose expectation is not constant or a deterministic
function of time implies the importance of variability. For a martingale (that
is, ignoring trend), there is a finite probability, which is a function of the
variability of the process, that the expectation of income at some future time
will be negative or zero, and that the firm will on average fail. The expectation
of all future incomes is changed with each observation. Hence, investors face
greater risk than under the other type of income process. The value of the
5. The varianceof income changesand higher moments are reduced,but the covarianceof successivechangesis reducedin the same order.Hence, the correlationcoefficientfor successivechanges
is roughly the same for the observed and unsmoothedseries. We observe essentially zero serial
correlation.
6. Of course, other types of smoothing can be investigated.In a paper which provides valuable
insights in other ways, Beaver [4] adopts a different definition of "smoothing."He hypothesizes
a non-manipulativetype of smoothing,differentfrom that which exists in the literature,and demonstrates its consistencywith our results.
7. Either process,when superimposedwith a trend, producesgrowth. A submartingaletherefore
can exhibiteither or both of two kinds of growth: on average,due to the expectationof trend; and
by chance,in spite of the zero expectationof non-trendchange.

This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

666

The Journal of Finance

firm should change in the order of a normal proportionality factor times the
change in income, reflecting the changed expected profitability.
The interpretationof decline and growth processes and of income variability
are therefore significantly different when income is generated by a submartingalemechanism.
Given that the nature of the income process has importance both for the
successful smoothing of income and for the interpretation of growth in the
incomes of firms, what is the prior evidence and what does it imply about the
income generating process?
IV.

PRIOR EVIDENCE

The original work of Little [28] and the later and larger study of Little and
Rayner [29] indicate that successive growth rates in the incomes of British
companies are random. Both studies use percentage changes in incomes to
measure growth rates and are concerned with the period-to-period stability
of those growth rates. As the authors recognize, the use of such period-toperiod percentage changes biases the results towards randomness.
Lintner and Glauber [27] also investigate the relationship between growth
rates in successive periods. However, the periods are longer, namely 5 and 10
years, and the sample consists of the 309 U.S. corporations on the Compustat
industrial tapes with positive dividends in each of the years 1946 to 1965
inclusive. Using log arithmetic data, growth is measured as the slope coefficient
in the regressions of various income variables on time over the 5 and 10
year periods.8Although they find very small cross-sectionalcorrelationbetween
the growth rates of successive periods, Lintner and Glauber are not prepared
to accept the hypothesis that the successive growth rates are independent:
"any conclusion to the effect that nothing but a table of random numbers is
relevant to growth in the real world itself would be premature and unwise."
[27, p. 8].
Brealey [6 and 7] follows Lintner and Glauber in studying the incomes of
U.S. corporations. However, he investigates changes in incomes instead of
growth rates in incomes. From cross-sectional correlations of changes in
incomes for various lags, runs tests and financial analysts' predictions based
on past accounting data only, Brealey effectively concludes that incomes follow a martingale [6, p. 13]. In his later book Brealey [7] reviews his own
study and that of Lintner and Glauber and does not change his conclusions.
Other evidence comes from less direct analyses. Fama and Babiak [17]
note in their dividend study that signs of earnings changes are nearly independent over time. Ball and Brown [2] find that changes in earnings capture the
new information which the stock market sees in an income number. Ball and
Brown [3] find that the assumption E(Yt) = Yt1, where Yt is the income
number in year t, leads to less error in measurementof the expectation of Y
than using an average of past Y's.9
8. The income variablesused by Lintner and Glauberare sales, operatingincome, earningsbefore
interest and taxes, aggregatedollar earnings,earningsper share and dividendsper share.
9. Specifically,a covariancebetween Y and another series is measuredwith less error in first
differences(which assumes E(Yt) = Yt-1) than in levels (which averages Y over the whole
time series).
This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

Propertiesof AccountingIncome

667

This evidence suggests that incomes conform to the specific kinds of nonrandomness which are implicitly assumed by using specific tests. In our
examination of the incomes of U.S. corporations, we do not impose such restrictive assumptions since it is difficult to hypothesize a specific form of
time series behavior for incomes. While the theory of efficient markets may
yield specific hypotheses for the time series behavior of market prices of
securities, there is no such theory for firms' incomes. The theory of the
firm is comparativestatic rather than stochastic, and the properties of accounting measurementrules are not well understood.Consequently,our investigation
is essentially a descriptive exercise and it uses a number of differenttests.
Since we are concerned with the income generating process per se, we do
not use the cross-sectionalapproach of Brealey, Little, and Little and Rayner.
Our analysis is deliberately confined to the expectations of the probability
distributions of change, ignoring other distributional features.10
V. DATA
Data are from Standard and Poor's Cornpustat file for the twenty years
1947-1966.11 Firms with less than twenty years of data are excluded from the
sample because the estimating procedures are sensitive to both few and
missing observations. As a consequence, the incomes of fewer than the approximately 900 firms on the S & P file are investigated, the number differing
according to the specific definition of net income which is used.
The effect of our sample selection procedure probably is to over-estimate
the importance of trends which are imposed upon expectations. Because the
S & P file contains only survived firms, because it contains only large firms
which are presumably older survivors than average, and because we do not
accept firms without early data and are presumably left with even older survivors than the S & P average, we probably have fewer average decreases in
income in our sample than have occurred over the time period under study.
Failed firms represent an obvious bias; age represents another. A proportion
of unexpected positive deviations from expectation will appear to constitute
an expectation of upward trend, since a proportion of offsetting chance
negative deviation will have been removed by the selection of only surviving
and older firms.'2
The importance of this upward bias in income relative to genuine expectations of trend (due, say, to reinvestment) cannot be determined within the
selected sample of firms. In an extension of this study we consider the effects
of these biases by investigating different samples of firms over different time
periods. Preliminary results indicate that the effects are minimal.
Four definitions of "income"in year t, Yt, are studied. They are:
10. For example we do not test directly whether incomes arise from a single constant process.
The more general form g(Yt.l I YO, . . . , Yt), where g is the distributionfunction for Ytl,:
is not considered.
11. Tape used is dated July 7, 1967.
12. Any competitivesituation with selectionof succeedingentities faces this same ex post sample
selection bias, unless recordsare available of failures after they drop out. The constructionof an
unbiasedsample of firms is thus a delicate procedure.We do not use log transformationsin this
paper becausewe do not wish to eliminatezero and negative incomes.
This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

668

Tke Journal of Finance

(1) Net income, after income taxes, as defined by Standard and Poor's.13
(2) Adjusted earnings per share, adjusted for stock splits and dividends.
(3) Net income, deflated by total assets, which might reduce reinvestment
effects.
Net
(4)
sales, which is chosen because of its possible dominance of the
income series, and because it could be less affected by income "smoothing" practices of accountants.
VI. TESTS AND RESULTS
As we suggested above, this study is a descriptive exercise since we do
not have theories of the firm or of the measurementof income. Further, due
to our limited number of observations for each firm, the results we obtain
may be sensitive to violations of the assumptions of each test. Analytical
results for most tests are for "large" samples."4We attempt to avoid both
issues by subjecting the income data to a number of different tests. Some of
their characteristics are noted below; conclusions are based upon their broad
tenor.15

Average changes-One indication of trend in a series is the number of increases relative to decreases [29, p. 390]. However, the test assumes symmetrically-distributedchanges. An alternative test for trend is to investigate
changes for the average firm; that is, to aggregate incomes over firms and then
to look at signs of changes in the aggregate series. The test is sensitive to
asymmetry, but the asymmetry possibly has been removed by the averaging
process.
Table 1 shows the time series of average net income and E.P.S. for 1947-65.
The averages are those used in [8]. Average net income is calculated as a
single average over 451 Cornpustat firms. Average E.P.S. is a weighted
average over the same firms, the weights being the adjusted number of common shares outstanding at the end of each year for each firm.
An ex post upward trend in both series is apparent. Unfortunately, the
sample cannot reveal the importanceof an ex post sample selection bias against
decreases in income, relative to an ex ante expectation of increase.
Runs tests-A "weak" test of independence is afforded by comparing the
actual and expected numbers of runs in a series. The test is "weak" because
it tests for independence conditional upon known probabilities of increases
and decreases in income, which must be estimated from the sample itself.
Thus it tends to fit the data too closely to the distribution which is assumed
to generate the observations.1 Furthermore, it tests a specific form of independence: the randomnessof the sequential arrangementin signs of deviations
within a finite series.
13. S & P exclude"extraordinaryitems" which are shown in companyreportsnet of tax. Where
gross figuresare available,extraordinaryitems are included.Using Moody's IndustrialManual as a
reference,it would appearthat S & P thereby include most extraordinaryitems.
14. Kendalland Stuart [26].
15. Bennett and Franklin [5, p. 688] also suggest this procedure.
16. Roberts [35], Chapter5, p. 24. Partly for this reason, the distributionof runs by length of
run is not considered.Anotherreasonfor the omissionis the sensitivity of this test to discontinuities
in a series.
This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

Properties of Accounting Income

669

TABLE 1
TIME SERIESOF INCOME
AVERAGED
OVER
451 CompustatFIRMS
Net Income
Year

Levels

1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965

($m)
11.17
14.43
13.12
17.27
15.96
15.32
16.93
18.18
23.72
24.57
24.63
21.27
25.26
25.16
25.59
28.66
32.17
37.08
42.47

E.P.S.
Changes
($m)
+3.26
-1.31
+4.15
-1.31
-0.64
+1.61
+1.25
+5.54
+0.85
+0.06
-3.36
+3.99
-0.10
+0.43
+3.07
+3.51
+4.91
+ 5.39

Increases
Decreases

13
5

Levels

Changes

($P.S.)
1.207
1.550
1.399
1.829
1.646
1.551
1.696
1.775
2.250
2.259
2.221
1.867
2.188
2.143
2.143
2.393
2.659
3.020
3.420

($P.S.)
+0.343
-0.151
+0.430
-0.183
-0.095
+0.145
+0.079
+0.475
+0.009
-0.038
-0.354
+0.321
-0.045
0.000
+0.250
+0.266
+0.361
+0.400
11
6

Source: Brown and Ball [8, p. 61]. The 451 firms were those which met the more stringentdata
requirementsof those authors.

Waldand Wolfowitz[38] calculatethe exact distributionof the numberof

runs,R, underthe assumptionof independence,and show that:
2N1N2
1AR

+1,

and
2

OR

2N1N2(2N1N2 - N)
NN2(N+1)

where -R and oR are the expectation and variance of R respectively, N1 is the

number of cases when Yt > Ya (for any Ya), N2 is the numberof cases when
Yt < Ya, and N = N1 + N2. Further assuming that both N1 and N2 are
"large," the statistic
R-tl

is normally distributed, with limiting distribution normal (0, 1). The normal

holds "closely"for N1, N2> 10 ([13], p. 289; [31], pp. 414approximation

16), whichrequiresat least fivemoreobservationsthanoursamplepossesses.17
17. An adjustmentfor the normal approximationto the binomial distribution [37, p. 280] is
omitted becausewe are describingthe data, not testing a specifichypothesis.

This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

670

The Journal of Finance

Results for Ya = Yt-1 (that is, for runs in signs of changes) are reported

in Table 2. The Table gives the deciles of the cross-sectional distribution of

TABLE 2
RUNS IN SIGNSOF INCOMECHANGES
Decile
Mean
Net Income
z value

.1

.2

.3

.4

-0.03

-1.37

-0.94

-0.49

-0.22

-0.12

0.34 0.50 0.74

1.20

-0.04

-1.22

-0.04

-0.61

-0.22

-0.13

0.25 0.39 0.72

1.01

.5

.6

.7

.8

.9

E.P.S.
z value

Net Income
Firms with more runs than
expected under independence
Firms with fewer runs than
expected under independence

Total runs in sample

Total expected runs,
assuming independence

E.P.S.

Number

per cent

Number

percent

348

48.7

326

48.0

366

51.3

353

52.0

714

100.0

679

100.0

6522

100.0

6338

99.8

6524

100.0

6350

100.0

Z scores for the 714 firms with net income data and the 679 firms with E.P.S.
data. It also gives two comparisons of the observed number of runs in the
series with the expected numberof runs under the assumptionof independence.
The runs tests reveal that, on average, changes in both net income and
E.P.S. are essentially independently distributed. The mean Z for both series
is extremely close to zero.'8 The actual number of runs is very close to the
expected number under the assumption of independence.
Serial correlation-The analytical serial covariance of changes in equallylagged drawings from an independently-distributedprocess is zero. The expectation of the computed serial correlationcoefficientof "large" samples from an
independent process also is zero. With "large" samples, the computed coefficient is insensitive to non-normality [26]. With "small" samples, and assuming normality,
1N 1'
and
2

08

T-

(T -1)2

where 9,, and a. are the expectation and variance of the computed coefficient S, N is the number of changes in the series, and T is N less the size
18. The median Z is sensitiveto the fact that R is an integer. However, it should also be noted
that there will be cross-sectionalcorrelationin firms'incomes.

This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

Propertiesof AccountingIncome

671

of the lag (N - 1 for successive differences). The distribution of S is approximately normal'9 and hence the statistic
Z

S
-

is approximately normal (0, 1).

Estimated coefficients are presented in Table 3 for net income and E.P.S.
with lag of one (that is, for successive first differences), and in Table 4 for
lags of two to five. The tables once again give cross-sectional data. Table 3
also gives the cross-sectional distribution of the standard normal deviate Z,
calculated on the assumptionof independence.
The serial correlation coefficients are in, agreement with the tenor of the
runs tests, indicating that net income changes are essentially independent and
that there is only a very low probability of a compensating mechanism for
E.P.S. Most mean and median coefficients are very close to the expectation
for an independent normal process. For example, with lag one, N = 19 and
assuming independence and normality, the expectation of the computed coefficient is -0.056 and its standard deviation is 0.228. The mean coefficient for
net income is -0.075, and for E.P.S. it is -0.198. Table 4 reveals that the
coefficients for lags two and five are even closer to their expectations than
those for lag one. The most extreme mean or median coefficient in either table
is -0.200. This is not only an extreme observation; it also implies a mere
4 per cent explanatory power for an autoregressiveprediction model.20
Mean squared successive difference-Runs tests give some indication of
the source of any nonrandomnessin a series. An excess of actual over expected
runs suggests a compensating mechanism, while a deficit is evidence of either
persistence, or of a mixture of processes with alternating parameter values.2'
The serial correlation coefficient has the added advantage of being sensitive
to the size of successive differencesand not merely their sign. The ratio of the
mean squared successive difference to an estimate of the variance possesses
some of the characteristics of both of these tests.
Define
N-1
=2

Z, (YVt1

t=l

- YJ)2/2(N-

1),

where Y't are successive differences. Then 82 is an unbiased estimate of the

variance of Y', if Y' is identically and independently distributed with finite
variance.
Define
N

t=1

19. Anderson[1, p. 7] and Malinvaud[30, pp. 292-94].

20. The analyticalserial correlationcoefficientfor first differencesin a stable stationary process
is -0.5. The computedcoefficientsclearly are of a lessermagnitude.
21. See Roberts [35, pp. 5-13].
This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

672

The Journalof Finance

z
Net
E.PDS.
value
value
Income
Coefficent
CoeffiCient
z

-0.42
Mean
-0.200 +0.01
-0.030

.1
-1.74
-0.453 -1.45
-0.386

.2
-1.38
-0.371 -1.01
-0.286

DISTRIBUTION
OF
SERIAL

.3
-1.10
-0.307 -0.78
-0.233

.4
-0.41
-0.84
-0.247 -0.150

CORRELATION
Lag
TABLE
I
3

COEFFICIENTS,
.5 Decile NET
-0.62
-0.198 -0.08
-0.075
INCOME
-0.42
-0.151 0.29
0.009

.6
CHANGES

-0.15
0.62
-0.090 0.085

0.2
0.09
-0.036 1.18
13

.7

.8

.9
0.49
0.057 1.94
0.388
This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

Properties of Accounting Income

Net
E.P.S.

673

LagE.P.S.

Net
Lag

Income
Expressed
as

Income
5 4 3 2

5 4 3 2

Mean

standard
Mean
+0.366
-0.584
normal

-.061
+.023
-.076 +.055
-.007
+.006
-.040
+.010

deviates.

.1

-1.91
-1.37

.1
RATIO
OF

-.331
-.375 -.277
-.300
-.306
-.368
-.346
-.320

DISTRIBUTIO
OF

.2

-0.86
-1.44

MSSD

.2

SERIAL

-.231
-.163
-.255 -.165
-.221
-.150
-.202
-.275

TO
.3

-1.15
-0.48

.3
ESTIMATED

-.163
-.182 -.085
-.131
-.150
-.140
-.128
-.197

CORRELATIO
Lags
TABLE
2-5 4

TABLE
5

.4

-0.92
-0.13

.4
VARIANCE

-.131 -.008
-.088
-.115
-.079
-.068
-.129
-.070

COEFFICIEN
Decile
.5

Dedile

.5

-.052
-.073
-.081 +.047
-.024
+.001
-.013
-.067

0.22
-0.73

INCOME

(STANDARDIZED)a
.6

0.63
-0.45

-0.18
1.04.7

1.67.8
0.26

NET

+.031
-.009
-.027
+.053
-.030 +.118
-.002
+.050

+.111
+.069
+.035
+.096
+.020 +.193
+.087
+.120

+.132
+.159
+.098 +.262
+.200
+.205
+.184
+.209

.6
CHANGES

.7

.8

.9

2.36.9
0.84
+.318
+.319
+.321
+.313
+.315UTC
This content downloaded from 202.43.95.117
on +.259
Sat,+.208
10 Oct+.403
2015
08:01:49
All use subject to JSTOR Terms and Conditions

The Journalof Finance

674

where Y' is the average difference.Then s2 is the usual estimate of the variance
under the same assumptions. If the assumptions are violated, then the ratio
62/S2 measures the strength and sources of the violation. If, for example, a
persistent trend exists, then the ratio will be small, or if the series fluctuates
highly, then the ratio will be large. Hart and von Neumann show that under
the assumption of an identically and independently distributed process with
definedvariance, and for "large"N, E =2/s2 has
F= 1

N-

2
=

N2-1

and is distributed approximately normally.22Hence the statistic

Ce

is normal (0, 1).

Values for the standardized E value are reported in Table 5. Zero would
indicate an independent process, while positive and negative values would
indicate persistence and fluctuations, respectively. The results are, on the
whole, consistent with previous tests. The median estimated E for net income
is +0.22 standard deviations from its expected value, assuming independence,
and is -0.73 standard deviations from its conditional expectation for E.P.S.
Once again, independence of changes in income is suggested.
It is helpful to bear in mind that the computed serial correlation coefficients
do not constrain the average change to be zero, and therefore linearly detrend
the income series. Further, the MSSD tests are conducted on changes in
income, and the runs tests do not constrain the probabilities of increases and
decreases to be equal. Hence, each of these three tests is a different test of
the independence of deviations from a constant trend mechanism. Taken together with the observed trends in the aggregate time series of Table 1, the
tests consistently imply that net income and E.P.S. behave, on average, as a
submartingale.23

Partial Adjustment Models-We initially distinguished two extreme types

of processes: processes with means which are a deterministic function of time
and submartingales. The first type implies the independence of the expectation of income from the observed past incomes; in contrast, a submartingale
implies total dependence in the sense that the expected value becomes the
most immediate past observed income.
Partial adjustment models provide a natural method of determining the
extent of dependence of the expectation of income upon past income. Let
22. Hart and von Neumann [24, p. 211]. Note the close similarity with the familiar tests for
autocorrelationin regressionanalysis,where Y' is constrainedto equal zero.
23. Some of the tests assume normal distributions,in which case the implication would be a
random walk with an upward trend, i.e. a submartingalewith a normal distribution.The term
submartingaleincludesrandomwalks and martingales(see footnote 1).

This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

Propertiesof AccountingIncome

675

Yt representa forecast of income in period t. Then let Yt be a function of the

prior period's forecast Yt-1, and the observed income Yt-1:
Yt aYt_

(+ O

o)Yt-1,

0 <.a < 1

Our objective is to estimate a from the income data. The a which produces,
ex post and on average, the "best" forecast of income is deemed to characterize the ex ante degree of dependence of the expectation of income upon past
observed values. That is, we define:
Y0,..
to= E(Yt |YI

Yt-1)

where YOis the optimal forecast of Yt on average.24An a of unity implies

a martingale.An a of zero implies a process with a constant expectation. Intermediate values of a provide a linear measurement of the extent to which the
time series can be described by a process with a constant expectation.
Two complications immediately arise. First, there is the problem of specifying what a "best" average forecaster of income would be. We choose mean
absolute error as a measure of forecasting accuracy,
et I = I: Yt
E:I
t
t

y~t1,

in order to avoid assumptions concerning the distribution of the errors. However, mean absolute errorgives almost identical results to those of mean squared
error. (unreported)
Second, there is the problem of growth in incomes over time. The partialadjustment models described above do not perform well under the upward
trend of the past twenty years. For example, a process with an expectation
which is a strong positive function of time would be best fitted by the model
which gives the most weight to the most recent observation: by a martingale.
In order to accommodatetrends, and therefore to avoid biasing the estimate of
a upward, the range of smoothing models is generalized beyond the simple
partial-adjustmentmodels which assume that variables are unrelated to time.
As well as these "constant" models we investigate models which allow for
linear and quadratic dependence on time.25Detailed proofs and some explanation of the models are given by Muth [32] and Brown [9].
Values of a ranging between zero and one and with increments of 0.05 are
used in making forecasts of Yt. These 21 values of a, combined with the three
levels employed, give 63 forecasting models which are tested. Thus, 63 forecasts of income are generated for each firm and each year. For a equal to
zero or unity, the three broad models degenerate into a process with a constant
mean and a martingale, respectively, for both of which time is not an argumodelsused to calculatethe
upon the classof forecasting
24. Theeidentityis also conditional
variousYt.
25. Thereare reasonsfor confiningthe analysisto the threebroadclassesof constant,linear,
between
and quadratic
models:we doubtwhether,giventhe availabledata,one coulddifferentiate
to limitthe studyfrom
makeit desirable
a largenumberof models;andthe estimation
procedures
a computational
pointof view.
This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

The Journalof Finance

676

ment. Thus, we are left with 59 sub-models overall, although there are 21 submodels within each broad class.
Results from Partial Adjustment Models-Actual Data
Forecasts in a given year are based upon actual observations in the series
up to the previous year. The 1948 forecast is the actual value for 1947, since
that is the only prior observation available.26Forecasts for subsequent years
employ progressively more observations.
For each firm and each variable, the sub-models are then ranked on their
total absolute errors over the 19 forecasts. Two sets of rankings are computed
for each firm: an overall ranking of the 50 sub-models and individual rankings
of the 21 sub-models within each of the three broad classes of models. The
ranks of each sub-model are then summed over all firms, giving two sets of
sums of ranks for each of the four income variables. The final precedure is
to rank these sums, giving ranks of 1 through 59 overall and 1 through 21 for
each class of models.
TABLE 6
SUB-MODELS RANKED FIRST FOR EACH VARIABLE ON THE BASIS OF MINIMUM
OVERALL RANKS
OVER N FIRMS
RANKS

FOR EACH

Fix

ARE BASED

ON MEAN

ABSOLUTE

SUM OF

ERROR

Variable

Model

Firms

Net Income
Earnings Per Share
Deflated Net Income
Sales

Constant
Constant
Constant
Linear

1.00
0.95
0.90
0.55

714
679
669
690

The sub-model which ranks first on the basis of overall rankings is shown
in Table 6. Except for sales, the generating process which best fits the data
is close to a martingale. A martingale is indicated for net income, and a
slight degree of serial dependence in changes in E.P.S. and deflated net income is indicated. The most recent observation in the time series appears, if
we take the results at face value and concentrate on expected values, to contain 100 per cent, 95 per cent, and 90 per cent of all the predictive information in the series, for net income, E.P.S. and deflated net income respectively.27
The fact that models which incorporate growth are not selected, except for
sales, probably reflects a misspecification of the growth process. Evidence is
presented below that this interpretation is reasonable. These results are consistent with those for the average series (Table 1).
The rankings of the models change slightly when computed within each
group. Ranking models 1 through 59 attaches a greater penalty to large fore26. As discussedbelow, this "startingproblem"biases our estimateof a. This is met in two ways.
First, the simulationsreportedbelow assessthis bias to be minimal.Second,in an unreportedanalysis,
the forecasting models rank in essentially the same order over only the last 15 observations, giving
the low a models time to "settle down."
27. Since the best sales forecast model under this criterion falls with the linear class, there is no
similarly natural interpretation for the sales variable. Note, however, that the absolute weight given
by the linear model to the most recent observationis twice times the value of a for that model.

This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

Properties of Accounting Income

677

TABLE 7
SUB-MODELS RANKED FIRST IN EACH CLASS FOR EACH VARIABLE ON THE
BASIS OF MINIMUM SUM OF RANKS WITHIN EACH CLASS
RANKS FOR EACH FIRM ARE BASED ON MEAN ABSOLUTE ERROR

as

Variable

Constant

Linear

Quadratic

Firms

Net Income
Earnings Per Share
Deflated Net Income
Sales

0.95
0.95
0.85
1.00

1.00
1.00
1.00
0.60

1.00
1.00
1.00
0.35

714
679
669
690

cast errors than does within-class ranking of 1 through 21. Table 7 presents
the rankings on a within-group basis. Again, the extreme relative importance
of the most recent observation is indicated, except for sales. The best "linear"
and "quadratic" models for each of the income variables all have a's of
unity, and therefore collapse into the martingale model. The best constant
model for sales is a martingale, but middle-sized a's outperform the martingale in both the linear and quadratic models, as well as overall. Since a linear
model with a= .6 gives a weight of 1.2 to Yt1, sales might be similar to a
submartingale (with a trend). We present more evidence on this point later.
The partial-adjustment models confirm that net income is, on average, a
martingale or a similar process. However, since the analytical properties of
partial adjustment models are not well known, we cannot place full reliance
upon them. Further, the availability of relatively few observations for the
forecasts of early years biases the findings against low a's. The importance
of such a bias is difficult to treat analytically. We therefore adopt further
tests in the following section.
Results from Partial Adjustment Models-Simulated Data
Data are simulated for twenty observations and four different generating
processes:
(a) a constant process, with stationary expectation and error term distributed normally;
(b) a linear process, with linearly increasing expectation and error term
distributed normally;
(c) a specific type of martingale, namely a random walk process with
error distributed normally; and
(d) a specific type of submartingale,namely a randomwalk with normallydistributed error, and with a linear trend through time.28
28. The generatingequationsare:
(a) Yt=zt
(b) yt = 8(t-1)

+ 4zt

(c) yt=yt_l+4zt

(d) yt = yt-1 + 4zt + 1

where Zt is drawn from the standardizednormaldistribution,and t = 1, .

.. , 20.

The expectations,

This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

The Journal of Finance

678

TABLE 8
SUB-MODELS RANKED FIRST FOR EACH SIMULATED PROCESS ON THE BASIS
OF MINIMUM SUM OF OVERALL RANKS FOR N SIMULATED FIRMS
RANKS FOR EACH FIRM ARE BASED ON MEAN ABSOLUTE ERROR

SimulatedProcess

Model

Constant with Noise

Linear with Noise
Martingale
Submartingale with Trend

Constant
Linear
Constant
Constant

.20
.20
.90
.95

200
200
714
200

TABLE 9
SUB-MODELS RANKED FIRST IN EACH CLASS FOR EACH SIMULATED PROCESS
ON THE BASIS OF MINIMUM SUM OFRANKS WITHIN EACH CLASS
RANKS FOR EACH FIRM ARE BASED ON MEAN ABSOLUTE ERROR

SimulatedProcess

Constant

a's
Linear

Constant with Noise

Linear with Noise
Martingale
Submartingale with Trend

.20
.55
.9o
.95

.10
.20
1.00
1.00

Quadratic

.05

200
200
714
200

.15
1.00
1.00

Tables 8 and 9 contain the rankings of the models, with simulated firms
ranked overall and within groups respectively. The partial adjustment models
discriminate submartingales from other processes even with only twenty observations available. There is some bias against low a's, as expected, since
a is estimated at 0.20 for a simulated stable and constant process (that is,
with a = 0). Whether this constitutes an effective bias in favor of the a's
close to unity is another question. The high a of .55 which the constant model
gives to the process whose expectation is a deterministic function of time shows
that more recent observations capture trend information. This is removed by
the linear and quadratic models.
While these results establish the discriminating power of the techniques,
they give no appealing measure of that power, and they do not assist in solving the equation of unconsidered forms of generating functions.29Thus, we
compare the rankings given the various models under the four actual income
series and the four simulated series. Table 10 quantifies the comparisons by
reporting the computed matrix of rank order correlation coefficients between
the rankings of the models when fitted to first the actual and then the simulated series. Rankings are computed overall (1 through 59).
We draw the following conclusions from the table:
(1) Deflated net income and E.P.S. behave very similarly to the simulated
martingale series, and the correspondence between the rankings for
variance, and trends are roughly estimated from a random sample of fifty of the actual firms.
Specificationof the distributionof Zt implies (c) and (d) are random walks. Experimentingwith
several variancesfor the error terms reveals that the results are not sensitive to the size of the
variances.
29. See footnote 24.

This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

Propertiesof Accounting
Jncome

679

TABLE 10

RAxxN
ORDER
CORRELATION
COEFFICIENTS
BETWEEN
THERANKINGS
OFTHEFOURVARIABLES
ANDTHERANKINGS
OF FouR SIMULATED
BASEDOFMINIMUMSUM
PROCESSES
OF OVERALL
RANKS
ActualVariabIes

SimulatedProcesses

Variableor Process

N.L.

Net Income
Earnings Per Share

1.0000
.9210

1.0000

Deflated Net
Income

.9015

.99S2

Sales

.7528

.4798

.4327 1.0000

.2405
.7181

.2974
.7406

Constant with
Noise
Linear with Noise

-.0270
.5467

.P.S.

DJ.

1.0000
.6325

1.0000

.5389

.9802

S.T.

1.0000

-.5342
.0751

1.0000
.6135
.0932

Martingale

.9742

.9788

.9671

.6232

Submartingale with
Trend

.9950

.9272

.9063

[.7436

-.0586

1.0000

The boxed numbersrepresentthe highest correlationbetween the ranks of the particularvariable

and the simulatedprocesses.
The hypothesisthat there is no relationshipbetween the rankings of two variables cannot be
rejectedif Jrl> .22 at the 95 per cent level, and if frI> .302 at the 99 per cent level.

net income and the simulatedsubmartingalewith a linear trend is

amazinglyclose. The deflatedand undeflatedincomevariablesdo not
look like processeswith means which are deterministicfunctionsof
time.
(2) Sales correspondsmost closely to a submartingalewith a trend,as we
earlierspeculated.However,the associationis not as strong as that
for the other variables.One interpretationof the "low" rank order
correlationof 0.74 betweenthe rankingsgiven to the sales series and
the simulatedsubmartingalewith a trend is that partial-adjustment
models do not performwell for such series. Table 8 presentssome
evidencein this regard:the trend in a simulatedsubmartingalewith
a trendis not identifiedby the partial-adjustment
models.
(3) The actualnet incomedata give similarrankingsto both a martingale
and a submartingale
with a trend.This is mostlikely due to the higher
growthrate in net incomeand the consequentialselectionof high a
modelswhich we referredto earlier.
(4) The partial-adjustment
class of forecastingmodels differentiatesprocesses with expectationswhich are constantor deterministicfunctions
of time from submartingaleprocesses.Hence the simulationresults
add credenceto our earlier conclusions.Their discriminatorypower
is greatestin non-trendsituations,as evidencedby the rank ordercorrelationsof 0.09 and -0.06 betweenthe rankingsgiven to simulated
submartingaleprocessesand the simulatedstationaryconstantmean
process.
In general,the simulationssupportthe previoussection.In that section it
This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

680

The Journal of Finance

was concludedthat net incomeis best describedby a martingale.However,

that section did not fit submartingaleswith trends to the net income time
seriesand the simulationdata suggeststhat those processesmay also be good
descriptionsof net income.Consequently,our conclusionfrom both sections
is that incomecan be characterizedon averageas a submartingaleor some
similarprocess.The term submartingaleincludesthe martingaleprocess.
VII.

SUMMARYOF

RESULTS

The evidenceof independencein detrendedincomechangesis compelling.

Results from a variety of testing procedureslead us to the conclusionthat
measuredaccountingincomeis a submartingaleor some very similarprocess.
We do not investigateincomewhich is calculatedin ways which differ from
conventionalaccountingpractice. Further, our conclusionsare necessarily
baseduponmeanand medianresults (becauseof small-sampleproblems)and
do not investigatewhetherspecificfirmsare systematicoutliers.Subsequent
researchaddressesthe latterissue.
The conclusionthat measuredincomeis a submartingale
is consistentwith
the earlierresearchreportedat the beginningof this paper.This conclusion
has importantimplications,some of which are spelledout above.
REFERENCES
1. R. L. Anderson."Distributionof the Serial CorrelationCoefficient,"Annals of Mathematical
StatisticsXIII (1942), 207-14.
2. Ray Ball and Philip Brown. "An Empirical Evaluation of Accounting Income Numbers,"
Journal of AccountingResearchVI, No. 2 (Autumn, 1968), 159-78.
and
. "Portfolio Theory and Accounting,"Journal of Accounting
3.
ResearchVII, No. 2 (Autumn, 1969), 300-23.
4. WilliamH. Beaver. "Time SeriesBehavior of Earnings."Paper presentedat the Conferenceon
Empirical Research in Accounting, Graduate School of Business, University of Chicago,
May 21, 1970.
5. Carl A. Bennett and Normal L. Franklin.StatisticalAnalysis in Chemistryand the Chemical
Industry. New York: John Wiley & Sons, Inc., 1954.
6. Richard A. Brealey. "StatisticalProperties of Successive Changes in Earnings,"unpublished
paper,March, 1967.KeystoneCustodianFunds, p. 13.
. An Introduction to Risk and Return from Common Stocks. Boston: M.I.T.
7.
Press, 1969.
8. Philip Brown and Ray Ball. "Some PreliminaryFindings on the Association Between the
Earnings of a Firm, Its Industry, and the Economy,"Empirical Research in Accounting:
SelectedStudies, 1967. Supplementto Vol. V, Journal of AccountingResearch,55-77.
9. Robert G. Brown. Smoothing,Forecastingand Prediction of Discrete Time Series. Englewood
Cliffs,N.J.: Prentice-Hall,1962.
10. RonaldM. Copeland."IncomeSmoothing."EmpiricalResearchin Accounting:SelectedStudies,
1968. Supplementto Vol. VI, Journal of AccountingResearch,101-16.
and Ralph D. Licastro. "A Note on Income Smoothing,"Accounting Review
11.
XLIII, No. 3 (July, 1968), 540-45.
12. John Y. Craggand Burton Y. Malkiel. "The Consensusand Accuracyof Some Predictionsof
the Growthof CorporateEarnings,"Journal of FinanceXXIII, No. 1 (March, 1968), 67-84.
13. Wilfrid J. Dixon and Frank J. Massey, Jr. Introduction to Statistical Analysis. New York:
McGraw-Hill,1957, second edition.
14. Nicholas Dopuch and David Drake. "The Effect of AlternativeAccounting Rules for Non30. In a subsequentpaperWatts [39] investigatesthe incomesof 32 firms over the 38 year period
1927-1964.He uses an identificationand estimationtechniquewhich allows a much broaderrange
of processesthan the partial adjustmentmodels we use. For 6 of the 32 firms Watts is able to
reject at the .01 probabilitylevel the hypothesisthat the incomes of those firms are randomwalks
with or without trend.

This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions

Properties of Accounting Income

15.
16.
17.
18.
19.
20.
21.
22.
23.

24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.

681

subsidiaryInvestments."EmpiricalResearchin Accounting: Selected Studies, 1966. Supplement to Vol. IV, Journal of AccountingResearch,192-219.
Eugene F. Fama. "The Behavior of Stock Market Prices," Journal of Business XXXVIII
(January, 1965), 34-105.
. "EfficientCapitalMarkets:A Review of Theory and EmpiricalWork,"Journal
of Finance XXV, No. 2 (May, 1970), 383-417.
and Harvey Babiak. "DividendPolicy: An EmpiricalAnalysis,"Journal of the
AmericanStatistical AssociationLXIII (December, 1968), 1132-61.
, Lawrence Fisher, Michael C. Jensen, and Richard Roll. "The Adjustment of
Stock Prices to New Information,"InternationalEconomic Review X (February, 1969),
1-21.
LawrenceFisher. "Some New Stock Market Indices,"Journal of Business XXXIX (Supplement, 1966), 191-225.
Jean-Marie Gagnon. "PurchaseVersus Pooling of Interests: The Search for a Predictor,"
EmpiricalResearchin Accounting: Selected Studies, 1967. Supplementto Vol. IV, Journal
of AccountingResearch,187-204.
Myron J. Gordon. "Postulates,Principles and Research in Accounting,"Accounting Review
XXXIX, No. 2 (April, 1964), 251-263.
. Discussion of Dopuch and Drake [14]. Empirical Research in Accounting:
SelectedStudies, 1966. Supplementto Vol. IV, Journal of AccountingResearch,220-223.
, Bertrand N. Horwitz and Philip T. Meyers. "AccountingMeasurementsand
Normal Growth of the Firm," in ResearchAccountingMeasurement,Robert K. Jaedicke,
Yuji Ijiri and Oswald Nielsen (eds.). Evanston, Ill.: American Accounting Association,
1966.
B. I. Hart and John von Neumann. "Tabulation of the Probabilitiesof the Ratio of the
Mean Square Successive Difference to the Variance,"Annals of Mathematical Statistics
XIII (1942), 207-214.
Samuel R. Hepworth. "Smoothing Periodic Income," Accounting Review, XXVIII, No. 1
(January, 1953), 32-39.
Maurice G. Kendall and Alan Stuart. The Advanced Theory of Statistics. Vol. III. London:
CharlesGriffin& Co., 1966.
John Lintner and Robert Glauber."HiggledyPiggledy Growth in America?"Paper presented
to the Seminaron the Analysis of Security Prices, GraduateSchool of Business,University
of Chicago,May 11-12, 1967.
I. M. D. Little. "Higgledy Piggledy Growth," Institute of Statistics, Oxford XXIV, No. 4
(Nov., 1962).
and A. C. Rayner. Higgledy Piggledy Growth Again. Oxford, U. K.: Basil
Blackwell, 1966.
E. Malinvaud. Statistical Methods of Econometrics.Chicago, Illinois: Rand McNally, 1966.
Alexander M. Mood and Franklin A. Graybill. Introduction to the Theory of Statistics.
New York: McGraw-Hill,1963.
J. F. Muth. "OptimalPropertiesof ExponentiallyWeighted Forecasts,"Journal of American
StatisticalAssociationLV, No. 290 (June, 1960), pp. 299-306.
P. D. Praetz. "AustralianShare Prices and the Random Walk Hypothesis."AustralianJournal
of StatisticsXI, No. 3 (November,1969), 123-139.
S. James Press. "SecurityPrices and the CompoundPoisson Process."Report No. 6707, Center
for MathematicalStudiesin Businessand Economics,GraduateSchool of Business,University
of Chicago,1967.
Harry V. Roberts. "StatisticalInference and Decision." Chicago, Illinois: The University of
Chicago,1966. (Mimeographed.)
Michael Schiff. "AccountingTactics and the Theory of the Firm," Journal of Accounting
ResearchIV, No. 1 (Spring, 1966), 62-67.
Robert Schlaiffer.Probabilityand Statistics for BusinessDecisions.New York: McGraw-Hill,
1959.
A. Wald and J. Wolfowitz."On a Test of WhetherTwo Samplesare from the Same Population," Annals of MathematicalStatisticsXI, 147-162.
Ross Watts. Appendix A to "The InformationalContent of Dividends," unpublishedpaper,
GraduateSchool of Business,University of Chicago, October, 1970.

This content downloaded from 202.43.95.117 on Sat, 10 Oct 2015 08:01:49 UTC
All use subject to JSTOR Terms and Conditions