Professional Documents
Culture Documents
Ohlson1980 PDF
Ohlson1980 PDF
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 support@jstor.org.
Wiley and Accounting Research Center, Booth School of Business, University of Chicago are collaborating
with JSTOR to digitize, preserve and extend access to Journal of Accounting Research.
http://www.jstor.org
1. Introduction
This paper presentssome empiricalresultsof a studypredicting
corporatefailureas evidencedby the eventofbankruptcy. Therehave
been a fairnumber of previousstudiesin this fieldofresearch;themore
notablepublished contributions are Beaver [1966; 1968a; 1968b],Altman
[1968; 1973],Altmanand Lorris[1976],Altmanand McGough[1974],
Altman,Haldeman,and Narayanan[1977],Deakin[1972],Libby[1975],
Blum [1974],Edmister[1972],Wilcox[1973],Moyer[1977],and Lev
[1971].Two unpublished papersby Whiteand Turnbull[1975a; 1975b]
and a paperby Santomeroand Vinso[1977]are ofparticular interestas
theyappear to be the firststudies which logicallyand systematically
developprobabilistic estimatesoffailure.The presentstudyis similarto
thelatterstudies,in thatthemethodology is one ofmaximum likelihood
estimation oftheso-calledconditional logitmodel.
The data setusedin thisstudyis fromtheseventies(1970-76).I know
of onlythreecorporatefailureresearchstudieswhichhave examined
data fromthisperiod.One is a limitedstudyby Altmanand McGough
[1974]in whichonlyfailedfirmsweredrawnfromthe period1970-73
and onlyone typeofclassification error(misclassification offailedfirms)
was analyzed.Moyer [1977] consideredthe period 1965-75,but the
sampleofbankrupt firms was unusuallysmall(twenty-seven firms).The
* Associate Professor,Universityof California,Berkeley. I gratefullyacknowledgethe
financialsupportfromthe Wells Fargo Bank. My thanksare also due to R. Wagnerand R.
Benin, who providedable and valuable assistance in the course of the project.G. Feltham,
R. Hamilton, V. Anderson,W. Beaver, and R. Holland supplied valuable commentson
earlierversionsof this paper. [Accepted forpublicationMarch 1979.]
109
Chapter X . 0 2 2 1 1 0 0 6
Chapter XI . 1 4 14 20 18 14 14 85
Other or unknown . 0 0 5 6 0 1 2 14
Totals ... 1 6 21 27 19 15 16 105
Iq
LO~ LO~
ei
t
CI
oo C,- e
~~0 I~ ~ 0
I O
C c1LC ei
Cl1
Cl1
o t- C
o o
Co1
LO
.4 I-1 CO
Ci2 C i2
* 0
4 0
00
-
Econ ; ; n Dz *
o E
can be quite long. In fact,for the group of eighteen firms,the lag was
always longerin the year whichwas closer to bankruptcy.The following
example is reasonably representative.Hers Apparel Ind. filedforbank-
ruptcy May 31, 1974; the accountants' report for the fiscal year-end
February28, 1974 is dated July19, 1974. In the previousyear,the report
was dated April 24, 1973. Note that the lead time between fiscal year
date of last "relevant" reportand date of bankruptcyis approximately
thirteenmonthsin this case (i.e., April24, 1973 to May 31, 1974). In 1974,
it took approximatelyfourand one-halfmonthsto complete the audit.
(I found many cases which exceeded fourand one-halfmonths.) There
were also a numberof firmsforwhich additional relevantreportscould
have existed. Under such circumstances,search procedures were at-
tempted,but with littlesuccess. For most of these cases, it appears as if
the firmssimplynever filedany additional reportswith the S.E.C. This
is by no means implausible,since firmscan apply forextensionof their
deadline, and afterbankruptcyhas actually occurredthere may simply
be no point in goingthroughan audit and preparinga standard annual
report.Of course, it is also possible that additional reportsdid exist,but
never got to the StanfordUniversityLibrary.In orderto play it "safe,"
I decided that no firmwas to be deleted because reportswere potentially
missing.As a consequence, any evaluation of a model based on this data
set probablyunderstatesthe predictiveclassificationperformance.
A sample ofnonbankruptfirmswas obtainedfromthe Compustattape.
Ideally, all reports for all firmssatisfyingthe population constraints
should have been includedas a controlsample. However,thiswas deemed
to be too costly and impractical (due to core memory constraints).I
decided instead that everyfirmon the Compustat tape (excludingutili-
ties, etc.) should contributewithonly one vectorof data points;the year
of any givenfirm'sreportwas obtained by randomprocedure.This led to
2,058 vectors of data points fornonbankruptfirms.The breakdowninto
exchange listings was as follows: New York Stock Exchange = 42%,
American Stock Exchange = 32%, Other = 26%.
and ratios are not quite comparable with those of Beaver [1966], the
resultshere are quite similarto the profileshe presented[1966,p. 82]. It
should also be noted that the standard deviations of the predictors
(except for size) are larger for year-1 firms,compared to nonbankrupt
firms.These differencesare significantat a 5-percentlevel or better.
Hence, as discussed in Section 2, standard assumptions of MDA are
unlikelyto be valid.
Three sets of estimates were computed forthe logit model using the
predictorspreviouslydescribed.Model 1 predictsbankruptcywithinone
year; Model 2 predicts bankruptcywithin two years, given that the
company did not fail within the subsequent year; Model 3 predicts
bankruptcywithinone or two years.A summaryofthe resultsare shown
in table 4. This table indicates that all of the signswere as predictedfor
Model 1. Only three of the coefficients(WCTA, CLCA, and INTWO)
have t-statisticsless than two,so the othersare all statisticallysignificant
at a respectable level. This includesSIZE, whichhas a relativelylarge t-
statistic.An overall measure of goodness-of-fit is givenby the likelihood
ratio index. The index is similarto a R2 in the sense that it equals one in
case of a perfectfit,and zero ifthe estimatedcoefficients are zero." For
Model 1, the ratio is 84 percent,and this is significantat an extremely
low a-level. The statistic "Percent CorrectlyPredicted" equals 96.12
percent;it is tabulated on the basis ofa cutoffpointof .5. That is, classify
the company if and only if P(XI., /) > 0.5. Whetherthis is a "good" or
"bad" result is not easy to answer at this stage, so furtherdiscussion
regardingthe model's predictivepoweris postponeduntilthe nextsection.
At thispoint,we can note that ifall firmswere classifiedas nonbankrupt,
then 91.15 percent would be correctlyclassified (2,058/(105 + 2,058)).
Thus the marginal(unconditional,prior)probabilityof bankruptcyis an
importantquantity in the above type of statistic.Further,there is no
apparent reason why .5 is an appropriatecutoffpoint,since it presumes
implicitlythat the loss functionis symmetricacross the two types of
classificationerrors.
Table 5 shows the correlationcoefficientsof the estimationerrorsin
Model 1. The coefficients ofthe financialstate variables (variables 1-4 in
the table) are uncorrelated with those of the performancevariables
(variables 5-9). Hence, both sets ofvariables contributesignificantly and
independentlyof each other to the likelihood function.This strongly
supports the contentionthat both sets of variables are importantin
establishingthe predictiverelationship.
Models 2 and 3 have somewhatweakergoodness-of-fit statistics,which
The likelihoodratio index is definedto be:
1 - log likelihoodat convergence/loglikelihoodat zero.
The index will take on the value of one in case of a perfectfit,since log likelihood at
convergencethen equals zero. If there is no fit,then obviouslythe index equals zero. See
McFadden [1973] forfurtherdetails.
c"
~~ c~~~o ~
Cq r 't
oq
oo
l
CO LO
U0~I Ho H
-
C Cl~
--
~
lf c CiclC~
CI Cl1
Cq U-D 00 Cq
- cI
O lN 0II
UrCS o
~
Lo lC
00
Cli oq~ %
0 Cq
m m >1 > e
- l Lo lf
m C0.0 0C0 m u0
- S~~~~~~~~~~~~~~~~~f
o
q-c-i oio
oro Cli ro
li
CZ ~ ~ ~ ~ ~ l
n C
Z? LoI- -e ? ' X
Q0~~~~~~~~~~~~~
00 m~~~~~
wr
C KL LL O
E-4~~~~~~~~~~~~~-
> O. CS > t cs . .
0 Q
c0 0t .: e
cn CC cn CC cn CC ~ ~ * l C
c\)~00
Q0
cc cc
. .
~ cc
.
cc
~ cc
~ cc o
~ 0
C Z C CJ
Z C Cio
Z
TABLE 5
CorrelationMatrix of (Estimated) Estimation Errors in Model 1
SIZE TLTA WCTA CLCA OENEG NITA FUTL INTWO CHIN
SIZE ....... 1 -.28 * * * * * * *
TLTA ...... 1 .32 * -.49 * * * *
WCTA 1 .46 * * * * *
CLCA 1 * * * * *
OENEG 1 * * * *
NITA 1 -.41 .40 -.44
FUTL 1 * *
INTWO.... 1 -.32
CHIN 1
valueofcoefficient
* = Absolute lessthan.20.
traded.
13 It would have been preferableto obtain data fornonbankrupt firmsby sampling10-K
reportsfromthe StanfordLibrary.However,this would have been a verycostlyprocedure.
6. Evaluation of PredictivePerformance
There is no way one can completelyorderthe predictivepowerof a set
of models used for predictive(decision) purposes. As a minimum,this
requires a complete specificationof the decision problem,includinga
preferencestructuredefinedover the appropriatestate-space. Previous
work in the area of bankruptcypredictionhas generallybeen based on
two highlyspecificand restrictiveassumptionswhen predictiveperfor-
mance is evaluated. First,a (mis)classificationmatrixis assumed to be an
adequate partition of the payoffstructure.Second, the two types of
classificationerrorshave an additive property,and the "best" model is
one which minimizes the sums of percentage errors. Both of these
assumptions are arbitrary,although it must be admitted that the first
assumptionis of some value if one is to describeat least one implication
of using a model. Much of this discussionwill thereforefocus on such a
(mis)classificationdescription.Nevertheless,the second assumptionwill
also be used at some points,since it would otherwisebe impossibleto
compare the resultshere withthose of previousstudies.The comparison
cannot be across models because the time periods,predictors,and data
sets are different.Rather, the question of interestis one of findingto
what extentthe resultsconformwith each other.
Withoutloss ofgenerality,one may regardthe estimatedprobabilityof
failure,PMX1,fi),as a signal which classifiesfirmi into one of the two
groups.Hence, it is ofinterestto describethe conditionaldistributionsof
these signals. Figure 1 shows the empiricalfrequencyofP(X,, /3)forthe
105 firms which failed within a year; fiis the vector of estimated
coefficientsobtained fromModel 1. Figure2 shows the frequencyforthe
2,058nonfailedfirmswhere/3is again takenfromModel 1. Figure3 shows
the probabilitiesfor 100 firmstwo periods priorto bankruptcy;the /P's
Occurrence (percentage)
10.0
8.0
60
.... .
40
.... .
0 20 30 4 90 0
0. . 5.0.0.0
... ... ... ... ~ ~ ~ ~ ~ ~ rooiltyofBnkupc
FI. .-irs
neyarprortobakupcy(15.irs
60
40
20
Probability of Bankruptcy
FIG. 2.-Nonbankrupt firms(2,058firms)
Occurrence (percentage)
2.0
1.0
I. .. .. :..:
::::: ::::
,,: ...:.......
. ...., ....m..
..'.
.. _
05.. . . . . . . . . ..
.. O..
... ..
Probabilityof Bankruptcy
FIG. 3.-Firms two years priorto bankruptcy(100 firms)
were taken fromModel 2. The mean probabilitiesare .39, .03, and .20,
respectively.This is, ofcourse,in accordance withwhat one would expect
on the basis of priorreasoning.
Using the data which underlie figures1-3, one can readily perform
analysis of classificationerrorsfordifferentcutoffpoints.The focuswill
be on a predictionof bankruptcywithinone year. A Type-I errorwill be
said to occur if P(X1, fi)is greaterthan the cutoffpoint and the firmi~s
nonbankrupt;in a similarfashion,one definesa Type-II errorforbank-
rupt firmsif the probabilityis less than the cutoffpoint. It would have
been preferableto performthe erroranalysis on a "fresh"data set and
thereby (in)validate the models estimnated.16 Due to the lack of data
beyond 1976,this was not possible at the time of the study.This should
not be a serious problem,however,forthe followingreasons.First,I have
'lbIt would, of course, have been possible to cut the sample in half and then go through
the usual kind of procedures. However, the primarypurpose of this paper is not one of
gettinga preciseevaluation ofa predictivemodel. Hence, it was decided thatthe fullsample
would be used in orderto produce the smallestpossible errorsofthe estimatedcoefficients.
Type I (percentage)
100r
80 - FREQUENCY OF ERROR
60
40
20
0
0 20 40 60 80 100
Type IT (percentage)
FIG. 4.
_-80
Type I Type II
60
40
20
60 40 20 0 20 40 60 80 100
of the misclassifiedfirms,I was unable to detect any ratio which was clearly out of line.
However,it is quite possible that a time-seriesanalysisofan extendedperiodwould indicate
thatsome ofthe firmshad "significant"above-averageoperatingrisks.By the use ofmarket
data, this problemwill be investigatedin the future.
TABLE 7
Type I-Type II Analysis forSelected CutoffPoints*
points too; that is, fora fixedlevel of one type of errorforboth models,
the complementaryerroris lower for Model 2. A close examinationof
table 7 will verifythis. To be sure,forsome errorrates Model 1 is better
than Model 2. It seems reasonable to suggest that the models are
essentiallyequivalent as predictivetools.
7. Conclusions
There are two conclusionswhich should be restated.First,the predic-
tive power of any model depends upon when the information(financial
report)is assumed to be available. Some previousstudies have not been
carefulin this regard.Second, the predictivepowersof linear transforms
of a vector of ratios seem to be robust across (large sample) estimation
procedures. Hence, more than anythingelse, significantimprovement
probablyrequires additional predictors.
REFERENCES
ALTMAN, E. "Financial Ratios, DiscriminantAnalysis and the Prediction of Corporate
Bankruptcy."Journal of Finance (September 1968).
. Corporate Bankruptcyin America. Lexington,Mass.: Heath Lexington,1971.
. "PredictingRailroad Bankruptciesin America." Bell Journal of Economics and
Management Science (Spring 1973).
, R. HALDEMAN, AND P. NARAYANAN. "ZETA Analysis:A New Model to Identify
BankruptcyRisk of Corporations."Journal ofBanking and Finance (June 1977).
, AND B. LORRIS. "A Financial Early WarningSystem forOver-the-Counter Broker
Dealers." Journal of Finance (September 1976).