Accounting Research Center, Booth School of Business, University of Chicago

A Five-State Financial Distress Prediction Model

Author(s): Amy Hing-Ling Lau
Source: Journal of Accounting Research, Vol. 25, No. 1 (Spring, 1987), pp. 127-138
Published by: Wiley on behalf of Accounting Research Center, Booth School of Business,
University of Chicago
Stable URL: http://www.jstor.org/stable/2491262
Accessed: 08-06-2015 17:52 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].

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

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
Journal of Accounting Research
Vol. 25 No. 1 Spring 1987
Printed in U.S.A.

A Five-State Financial Distress

Prediction Model
AMY HING-LING LAU*

1. Introduction
The modelpresentedhere extends previouscorporatefailureprediction
models (e.g., Beaver [1966], Altman [1968], Altman, Haldeman, and
Narayanan [1977], Ohlson [1980], Zavgren[1985], etc.) in two ways: (1)
instead of the conventional failing/nonfailing dichotomy, five financial
states are used to approximate the continuum of corporate financial
health; and (2) instead of "classifying"a firm into a certain financial
state, the new model will estimate the probabilitiesthat a firm will enter
each of the five financial states. The ranked probabilityscoring rule is
then used to evaluate the quality of such probabilisticpredictions. The
first extension enables the predictionof prefailuredistress in addition to
ultimate failure. The second extension conforms with more recent ad-
vances in predictionmethodologies(see, e.g., Epstein [1969] and Murphy
and Winkler [1970]). Together, the two extensions provide a better
approximationto the continuum of alternative financial judgment and
actions in reality.
Sections 2 and 3 explain the structure of the model and the data;
section 4 explainsthe statistical methodology(multivariatelogit analysis)
and presents the constructed models. These models are evaluated in
section 5.
2. Categorizing and Sampling of Firms
2.1 THE FINANCIAL STATES OF FIRMS
The five financial states used in this study are: state 0: financial
stability; state 1: omitting or reducingdividendpayments; state 2: tech-
* Oklahoma State University. [Accepted for publication June 1986.]

127
Copyright C),Institute of Professional Accounting 1987

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
128 JOURNAL OF ACCOUNTING RESEARCH, SPRING 1987
nical default and default on loan payments; state 3: protection under
Chapter X or XI of the Bankruptcy Act; and state 4: bankruptcyand
liquidation. States 1 to 4 are states of increasing severity of financial
distress.
Although a financially stable firm may reduce or omit dividends, say,
to finance capital investments, empirical studies by, e.g., Donaldson
[1969], Pettit [1972], Dielman and Oppenheimer [1984], and Gentry,
Newbold, and Whitford [1985] have shown that a firm that reduces
dividends is typically encountering some financial distress. Therefore,
this study uses "dividendomission or reduction"to representa financial
condition between states 0 and 2.
2.2 SAMPLE DESIGN
The prediction models are constructedwith an "originalsample"and
then tested with a "holdoutsample."Each sample contains 350 firms in
the financially healthy state 0; and 20, 15, 10, and 5 firms in states 1, 2,
3, and 4 respectively.These firms were selected as follows.
1. State 0 Firms. From the Compustat tapes, 350 firms which were
financiallyhealthy in 1976 (1977) were selected for the original (holdout)
sample. Every firm met the following conditions: (i) its asset-size fell in
the same range ($1.6 million to $120 million) as that of the financially
distressed firms; (ii) its 1971-77 financial reports were available; and
(iii) it experienced no financial distress or financial loss between 1972
and 1977.
2. State 1 Firms. The Compustat tapes were used to generate a list of
firms that reducedtheir annual dividend rate per share in 1976 (for the
original sample) or 1977 (for the holdout sample) by more than 40%
below that of the previousyear. From this list, 20 firms were selected for
each sample. Each of those selected had (i) a stable dividend payment
recordduring1972-75 and (ii) a decliningtrend in payments from 1976-
78. Of these 40 firms, only four (10%)subsequentlydeterioratedfurther
and ended up in more serious states of financial distress (i.e., states 2
and 3).
3. State 2 Firms. The Wall Street Journal Index (WSJI) and the
Standard & Poor Stock Reports were used to compile a list of firms that
had either filed for protection under ChapterX/XI during 1977 to 1980
or had C-ratedbonds. The SEC 10-K reportsof each of these firms were
examined to identify those that defaulted loan interest and/or principal
payments during 1976 and 1977. Fifteen loan-defaulting firms were so
obtained for each of the two samples. For such firms included in the
original sample, 12 (80%)eventually filed for ChapterX and XI protec-
tion, whereas among such firms included in the holdout sample, only 7
(46.7%)did so.
4. State 3 and 4 Firms. From the WSJI list of bankruptand Chapter-
X/XI firms, for each sample, 10 Chapter-X/XI and 5 bankrupt firms
that had publicly available 10-K reportswere selected.

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
 FINANCIAL DISTRESS PREDICTION MODEL 129

Zmijewski's [1984] study indicates that while the above procedure

oversamplesdistressingfirms and favors firms with completedata, these
biases will not materiallyaffect the resultant models'predictivevalidity.

2.3 PREDICTION HORIZONS

Predictive models with three different prediction horizons are con-
structed. That is, "year-i"predictionmodels are constructedwith 1974/
75 financial informationto predict financial distress in 1976, and simi-
larly "year-2"and "year-3"models are constructed with 1973/74 and
1972/73 financial information respectively to predict 1976's financial
distress. The holdout sample is used to test these models' ability to
predict 1977 distress.

3. The Explanatory Variables

3.1 VARIABLE SELECTION
This study uses Donaldson's [1969] (see also Heath [1978] and FASB
[1980]) "financial flexibility" concept as a frameworkfor selecting the
explanatory variables. These variables are summarizedin table 1 and
explainedbelow (see Lau [1982] for more detailedjustifications).
The "financial flexibility"concept views the maintenance of a firm's
funds-flow balance as the essence of its solvency; during funds-flow
deficits, a firm's ability to recover depends on its "financial flexibility"
resources. Donaldson [1969] identified five "financial flexibility" re-
sources: borrowingcapacity, stock flexibility, cost flexibility, dividend
flexibility, and asset disposability. Variables X1 to X3 in table 1 are
intended to measure borrowingcapacity, and each of the four variables
X4 to X7 is intended to measureone of the four flexibility resources.
The "trend variables"in group 2 of table 1 are used because of (a)
earlier studies (e.g., Donaldson [1969]) indicating that a firm tends to
reduce its capital expenditurewhen financial difficulty approaches,and
(b) the well-recognizedclose relation between a firm's funds-flowand its
solvency (see, e.g., Heath [1978] and Emery and Cogger[1982]).
X10is intended to capture the data that a firm currently in financial
distress is more likely to be in greater distress later than a currently
healthy firm. This variable differs from X6 since it indicates whether a
distress action has been taken, whereas X6 indicates whether a financial
flexibility resourceexists.

3.2 PRELIMINARY DATA ANALYSIS

The value of each explanatoryvariableis obtained for each of the five
groupsof firms (correspondingto the five different states) and for each
of the three different prediction horizons (i.e., for the year-1, -2, and -3
models). Table 2 gives the means, medians, and standard deviations of
the values for the year-1 models. For each variable, the nonparametric

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
130 AMY HING-LING LAU

TABLE 1
Summary of Explanatory Variables Used in Five-State Financial Distress Prediction
Models

Variable Brief Definitionof Variable Abbre- Natureof

Number viation the Variable
Group 1: Financial Flexibility Variables
Xi = 1 if one of the firm's loan agreements contains 3 LRT dichotomous
or more restrictive terms and the loan's interest is
above the prime rate.
= 0 otherwise.
X2 (Firm's Debt-Equity Ratio) + (Industry Debt-Equity DER ratio
Ratio).
X3 Working Capital Flow/Total Debt. WFTD ratio
X4 Stock Price Trend = TCSP ratio

(Ht - Ht1) + (Lt -Lt-)

Ht + Ht-1 + Lt + Lt-1
where Ht and Lt are, respectively, the high and low
values of the range of stock prices in year t.
X5 (Firm's Operating Expense to Sales Ratio) *. (Indus- OPES ratio
try's Operating Expense to Sales Ratio).
X6 = 1 if no dividend is being paid currently, DCSD dichotomous
= 0 otherwise.
X7 = 1 if the firm liquidates its operating assets in the LOPA dichotomous
period and there is no decreasing trend of earn-
ings flow,
= 0 otherwise.

Group 2: Two Trend Variables

X8 Trend of Capital Expenditures = TCEP ratio

(Ki -Ki-)
(Ki + Ki-1 + Ki-2 + K_3)/4

where Ki = capital expenditure in year i.

Xg Working-Capital Flow Trend = TWF ratio

(WFt - WFj1)
(WFt + WFt-1 + WFt-2 + WFt-3)/4

where WFt = working capital in year t.

Group 3: Indicator of Current Financial State
X10= 1 if dividend payments are omitted or reduced DVD dichotomous
more than 40% in the period,
= 0 otherwise.

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
 FINANCIAL DISTRESS PREDICTION MODEL 131

2 1

State
State
State
'VariableState StateYear-1
State X1-LRT
X2-DER
DVDTWF LOPA
TCEP DCSD
OPES WFTD
TCSP DER
LRT 4: 3: 2: 1: 0: Xg-TWF
X1o-DVD X7-LOPA
X8-TCEP X4-TCSP
X5-OPES
X6-DCSD X3-WFTD Variables3
= == = = == = = =
model:
trend
trend trend loan
omitting
financial
technical 1.16 Mean
of of
omission of industry
industry protection model
0.011 -0.006
0.024 0.171
0.097 0.035
0.359 0.011
0.964
or
bankruptcy
or liquidation
distribution 0
of of and with
capital restrictivedefault
under 0 0 0
definitions-five-state 0
working-capital stability; (n=
common definitions-five-state
reducing
and 0.208
-0.024 0.233
0.024 0.855 Median
normalized
normalized
flow 1.0265 350)
reduction
common
operating to terms; Chapter prediction
stock
of working-capital X defaultfinancial
liquidation.
Group'
expenditure;
stock
flow;assets; or ondividend 0.106
total financial 0.093
0.719
0.296 0.699
0.377 0.119
0.349
0.618
0.106
STDV
prices;
operating
XI horizon
distress
dividend debt of loan of
debt-to-equity
distress 1 0 Means,
ratio; the payments; 0.65 0.35 1.073 0.15Mean
expenses
dividends;ratio; -0.186
-0.688 0.234
-0.0721.216
models:
year. 1
to
payments. models:payments; 0 0 0 (n
sales 1.0 = Medians,
Bankruptcy -0.553 0.916
-0.039 -0.0730.872 Median
0.122 20)
and
ratio; 0
Act;
0.489 0.423
0.379 0.489 0.859
0.092 1.325
0.412 0.366
STDV
Standard
0.20.8 -0.11
0.267
-0.356 -0.0954.753
-0.307 1.327 Mean
0.867
2
0 0 (n Firms TABLE
1.0 1.0 = in Deviations
2
-0.52
-0.329 1.338 -0.07
2.65
-0.125 Median
15) for
State2
0.33
0.387 0.414
0.458 0.414
0.432 0.153
0.354 0.352
4.468 STDV Year-i1

0.2 0.50.9 0.8 Mean

12.04
-0.229
-0.606 2.252
-0.279
-0.148
3
0 (n=
0.51.0 1.0
8.43 Model-Original
-0.241
-0.427 1.846 -0.075
-0.286 Median
10)

0.55 10.94 STDV Sample

0.422
0.312 0.527 1.953
0.316 0.186 0.422
0.169

0.2 0.60.8 1.0

-0.311
-1.676 1.473-0.22516.09 Mean
-0.162
4
(n
0 1.01.0 =
1.529 -0.11 1.0 5)
-0.38
-0.299 -0.26517.48 Median
0
0.447
0.474
2.986
0.548 0.621
0.477 0.251
0.332 6.174 STDV

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
132 AMY HING-LING LAU

TABLE 3
Probability of Error in Rejecting Ho in Favor of H1: A Variable's Values for Five States
Come from Five Ranked Populations (Jonckheere's [1954] Test-Original Sample)
PredictionHorizon
Variables'
Year-1Model Year-2Model Year-3Model
X1-LRT 0.042 0.012 0.000
X2-DER 0.000 0.000 0.000
X3-FFTD 0.000 0.000 0.000
X4-TCSP 0.000 0.039 0.368
X5-OPES 0.053 0.103 0.078
X6-DCSD 0.034 0.000 0.008
X7-LOPA 0.000 0.000 0.000
X8-TCEP 0.000 0.025 0.053
X9-TFF 0.000 0.012 0.301
X1o-DVD 0.002 0.000 0.000
Variabledefinitions:
LRT = loan restrictiveterms;
DER = industrynormalizeddebt-to-equityratio;
WFTD = working-capitalflow to total debt ratio;
TCSP = trendof commonstock prices;
OPES = industrynormalizedoperatingexpensesto sales ratio;
DCSD = distributionof commonstock dividends;
LOPA = liquidationof operatingassets;
TCEP = trendof capitalexpenditure;
TWF = trendof working-capitalflow;
D VD = omissionor reductionof dividendpayments.

Jonckheere's [1954] test1 was used to test whether its values for the five
groups of firms come from five ranked populations, and table 3 gives the
significance level at which the null hypothesis (of randomness) can be
rejected in favor of the ranked population alternative. These results
strongly suggest that each of the ten explanatory variables (excepting
perhaps X5) assumes different (and appropriately ranked) values for
firms in different states and is a potentially useful predictor. The results
also indirectly support the viability of the five-state representation of
corporate financial condition. These conclusions are further confirmed
by paired t-tests (results omitted here) indicating that, for each variable,
the five state-means differ from each other under paired comparison.
Subsequently, correlation analyses results (omitted here) indicate that
the explanatory variables are not highly correlated with each other, i.e.,
each variable is nonredundant and there is little or no multicollinearity.
4. The Logit Predictive Models
4.1 THE MLA METHODOLOGY: BRIEF EXPLANATION
The financial distress prediction models are constructed using "mul-
tinomial logit analysis" (hereafter MLA). Consider a problem in which
1 For observations on a variable x drawn from k populations with cumulative distributions
F1(x), F2(x), ..., Fk(x), Jonckheere's [1954] distribution-free test considers the alternative
hypothesis:
F (x) < F2(x) < ... < Fk(x)
against the null hypothesis of homogeneous cumulative distributions.

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
 FINANCIAL DISTRESS PREDICTION MODEL 133

all firms will enter one of J = 5 states. Each firm's destiny is predicted
by K = 10 explanatory variables, designated x1, x2, . . ., x10. Defining pj
as the probability that a given firm will eventually enter state j, the logit
model postulates that the pj's of the firm can be estimated as follows:
(i) compute Z1= bixi + bj2x2 + *.* + by,10x10
for each state j = 0 to 4, (1)
J
(ii) then pj = exp(Zj)/ exp(Zj). J
j=1
The coefficients bjkcan be considered as the effect of the kth explanatory
variable on a firm's probability of entering state j. Constructing an MLA
prediction model consists of estimating the b/4's using an empirical data
set containing the actual final state and the values of the explanatory
variables of N firms. More detailed treatments of MLA can be found in,
e.g., McFadden [1973] and Nerlove and Press [1973].
4.2 COEFFICIENTS OF THE LOGIT FUNCTIONS
The QUAIL program (Berkman et al. [1979]) was used to construct
three logit prediction models, one for each prediction horizon. Each
model has five logit functions, one for predicting each of the five states.
Table 4 contains the functions for the year-1 model; functions for the
year-2 and -3 models have similar structures and are detailed in Lau
[1982].
The expected sign of each coefficient in each logit function depends
on the effect that variable has on a firm's final state. For example, a
higher X3 (see table 1) means higher financial flexibility, hence higher
probability of entering state 0 and lower probability of entering state 4.
Therefore, X3's expected sign is positive in the state-0 logit functions
and negative in the state-4 logit functions. The entries in table 4 agree
with these expectations. Similarly, expected signs of most other explan-
atory variables generally find agreement in table 4. The most conspicuous
inconsistency pertains to the debt-equity ratio variable X2 (DER). The
expected sign for this variable is negative for states 0 and 1, and positive
for states 2, 3, and 4. The empirical signs shown in table 4 for X2 are
inconsistent with this expectation. It is relevant to note that similar
inconsistencies exist in Ohlson's [1980] study (see [1980, table 4]);
perhaps the most intriguing result appearing in both Ohlson's and the
current study is that, for some explanatory variables, the empirical signs
agree with the expected signs for models of some prediction time-horizons
but not for the others.
5. Evaluating the MLA Models
5.1EVALUATING CRITERION I: PROBABILISTIC PREDICTION
SCORES
One feature of the MLA models in this study is the generation of
probabilistic (as opposed to deterministic) predictions. Currently, the

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
 134 AMY HING-LING LAU

4 3 2 1 0
State
State State
State State
3State (n (n (n (n (n
4: 3: 2: 1: 0: 2Year-1 = = = = =
DVDTWF LOPA
TCEP DCSD
OPES WFTD
TCSP DER 'Variable
LRT States3
5)
'Estimated 15) 20)
= = = = = = = = = = t-statistics 10)
t-statistics t-statistics
t-statistics t-statistics
350)
model:
omitting
financial
technical logit
trend
trend trend loan bankruptcy
protection
or
of of of industry
industry model
omission and
or liquidation
distribution default
under -132.5 -137.8
of of stability;
with 90.43 89.11 90.76
capital restrictiveand
working-capital reducing 2.507 2.479 2.523 -2.456 -2.555 X1(LTR)
coefficients
definitions-five-state
common definitions-five-state for
normalized
normalized
flow Chapter
reduction stock
to terms;
liquidation.
X default prediction
common
operating
of working-capital or ondividend
financial
-1.17
financial equation -0.965-1.389 -1.3711.158
-1.166-1.629 1.474 1.643
2.028
X2(DER)
expenditure;
stock total XI
flow;assets; prices;
operating (1).
of loan horizon
dividend debt distress
debt-to-equity
the payments; of
distress Logit
ratio;
expenses one
dividends; ratio; models: -829.8 -835.3 -836.9 1250.0 1252.0
to payments;
payments. models: year. -2.492 -2.497 -2.502 2.494 2.499 X3(WFTD)
sales Bankruptcy
Act; Coefficients'
ratio;
110.1 113.8
-72.79 -78.55 -72.56
-2.349 -2.485 -2.291 2.342 2.421 X4(TSCP)
Estimates'
for

77.15 -2.281 11.64 2.55

14.55 TABLE
-2.333
-9.1332.526 -8.7882.041 X5(OPES)
Year-1
4

Variables4

.0112.95 -37.6 -12.51

17.77 19.39
0.029 0.032 -0.021 -0.021 X6(DSCD)

Model2-Original

-115.1 -114.7
77.41 2.46
75.24 -2.502 -2.494 X7(LOPA)
1.452 -2.146
-8.269 Sample

42.1
42.62 -63.85-2.462
41.08 2.505 2.531 -61.95
2.456 -2.538 X8(TCEP)

197.8 194.2 -295.0 -289.5

2.494 2.539 2.489 -2.532 -2.486 Xg(TWF)
192.499

17.6 12.93
-1.761
-11.731 -11.02 -7.78
-1.765 -1.257 1.916 1.409 X1o(DVD)

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
 FINANCIAL DISTRESS PREDICTION MODEL 135
most "theoretically correct" approach to evaluate these probabilistic
predictions is the "ranked probability scoring rule" developed by Epstein
[1969] and Murphy [1971]. Given a probabilistic prediction:
F = (fl, f2, f&, f4, f5),
where ft is the predicted probability of state i occurring, if "k" is the
actual state eventually entered into, then the probabilistic prediction
score "S" for this prediction under the ranked probability scoring rule
is:
2 n \21
3 1
S =2 --
n-i [/
I( f3) +
I (Z f)I
/
2(n - 1) i=L =1 \(=i+i

i(2) k
For each prediction, the maximum possible score of 1 is earned when thefi.
a state
actual tr outouto
l st turns t bebK, and the probabilistic prediction is:
F = {fK= 1; and f, = 0 for i $ K}.
For example, consider two probabilistic predictions FA = (0.5, 0, 0.5, 0,
0) and FB = (0.5, 0, 0, 0, 0.5). If the actual state is state 1, a scoring rule
that ignores the ranking of states will give FA and FB the same score.
However, if state 1 is considered "closer" to state 3 than state 5 (i.e., the
states are "ranked"), then FA with f3 = 0.5 is better than FB with f3 = 0
and f5 = 0.5. Equation (2) reflects this ranking since it gives S = 0.875
for FA and S = 0.75 for FB. The probabilistic predictions of the MLA
models must be evaluated by a "ranked" scoring rule since the five
financial states are clearly ranked.
Applying a probabilistic prediction model to a group of n firms gives n
probabilistic prediction scores, and the prediction performance is repre-
sented by the sum of these n scores (SSn) as well as the ratio SSn/n
(since n is the maximum possible sum of scores). Table 5 gives the SSn
for each of the five groups of firms and for the entire set of 400 firms.
For example, it indicates that the probabilistic predictions produced for
the 15 state-2 firms in the original sample by the year-1 MLA prediction
model earned a total of 14.38 out of a maximum possible score of 15.
Except for the state-4 firms2 in the holdout sample, table 5 indicates that
the score earned by each group of firms is close to the maximum possible
score. For comparison, multiple discriminant analysis (MDA) was ap-
plied to the same data set, and the SSn's for the entire set of 400 firms
are given in the last column of table 5. It can be seen that MLA
outperforms MDA in every case, with larger differences in the holdout
sample.
2
Investigationof the poor performancefor state-4 firms in the holdout sample reveals
that two of the sampled "bankrupt"firms were actually financially healthy firms that
liquidatedfor the sole purpose of enhancing their shareholders'financial position. The
MLApredictionmodelperformedcorrectlyand identifiedthem as state-Ofirms.

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
136 AMY HING-LING LAU

TABLE 5
Aggregate Probabilistic Prediction Scores Earned for Different Groups of Firms
MultinomialLogit Models Multiple
Prediction Firmsin State' All Discriminant
Horizon Firms Models
0 1 2 3 4 (n = All Firms
(n = 350) (n = 20) (n = 15) (n = 10) (n = 5) 400) (n = 400)

Original Sample2
Year-1 Model 349.2 18.67 14.38 9.45 4.64 396.3 391.3
Year-2 Model 347.7 17.12 13.21 7.47 5.00 390.5 385.8
Year-3 Model 347.1 16.48 12.53 7.69 4.34 388.1 379.6

Holdout Sample3
Year-1 Model 336.0 17.33 13.25 7.80 1.74 376.1 332.1
Year-2 Model 335.2 16.87 11.30 7.82 2.52 373.7 369.1
Year-3 Model 334.9 16.52 12.35 7.52 2.84 374.2 365.7
1
State definitions-five-state financialdistressmodels:
State 0: financialstability;
State 1: omittingor reducingdividendpayments;
State 2: technicaldefaultand defaulton loan payments;
State 3: protectionunderChapterX or XI of the BankruptcyAct;
State 4: bankruptcyand liquidation.
2 Originalsample:financialinformationfrom 1972-75 is used to predictfirmsin financialdistressin
1976.
'Holdout sample:financialinformationfrom 1973-76 is used to predictfirmsin financialdistressin
1977.

5.2 EVALUATING CRITERION II: CLASSIFICATION ERROR

Because earlier models generated deterministic predictions and used

the performance criterion "the number of firms classified correctly," the
probabilistic predictions in this study are converted to deterministic ones
to facilitate comparison. Using the MLA-generated probabilities of a firm
entering each of the five possible states, the firm is "classified" into the
state j with the highest predicted probability of entering. A firm is
"correctly classified" if the firm actually entered j, otherwise the firm is
"misclassified." The percentages of firms correctly classified by the MLA
models under this "conversion" procedure are given in table 6, together
with the corresponding figures from seven related, earlier studies.
Three factors should be kept in mind when examining table 6. First,
firms in all the earlier models can enter only two states, but firms in this
study can enter five different states, implying more ways to misclassify
a firm. Second, in the earlier two-state prediction models, the model
typically computes a score for each firm. A firm is then classified into
one state or the other depending on whether the computed score is above
or below some cutoff point that is determined by minimizing the sample
classification error. A similar optimization procedure is not possible for
the five-state probabilistic prediction models. Third, in the three earlier
studies that performed validation tests (Deakin [1972], Blum [1974], and
Zmijewski [1984]), holdout samples were drawn from the same time
period as the original samples. Joy and Tollefson [1975] have pointed
out that such a procedure is not an intertemporal validation, and it does

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
 FINANCIAL DISTRESS PREDICTION MODEL 137
TABLE 6
Comparison of Percentages of Correctly Classified Firms

OriginalSample HoldoutSample
CorporateFailure
PredictionStudies Year-1 Year-2 Year-3 Year-1 Year-2 Year-3
Model Model Model Model Model Model
Beaver [1966] 87% 79% 77% Not Considered

Altman [1968] 95% 72% 48% Not Considered

Deakin [1972] 97% 95% 95% 78% 94% 88%

Blum [1974] 95% 90% 82% 95% 80% 70%

Altman, Haldeman,
and Narayanan 93% Not Considered 91% 89% 84%
[1977]

Ohlson [1980] 96% 95% Not Not Considered

Zmijewski [1984] 97.1% Not Considered 97.4% Not

Considered

This Study-Five-State
Model

Overall 96% 92% 90% 80% 79% 85%

State O0 99.4% 98.9% 99.1% 85.4% 86.5% 93.7%

State 1 65.0% 15.0% 10.0% 50.0% 20.0% 10.0%
State 2 86.7% 66.7% 46.7% 66.7% 33.3% 46.7%
State 3 70.0% 40.0% 30.0% 20.0% 10.0% 20.0%
State 4 60.0% 100.0% 80.0% 20.0% 20.0% 20.0%
1
State definitions-five-state financialdistressmodels:
State 0: financialstability;
State 1: omittingor reducingdividendpayments;
State 2: technicaldefaultand defaulton loan payments;
State 3: protectionunderChapterX or XI of the BankruptcyAct;
State 4: bankruptcyand liquidation.

not validate a model's prediction ability. The holdout and original sam-
ples used in this study relate to firms becoming financially distressed in
two different years. Therefore, the validation in this study provides at
least some confirmationof the MLA-model'sprediction ability.
6. Concluding Summary
This study has demonstrated the feasibility of constructing MLA
models that compute the probability that a firm will enter each of the
five different financial states. The probabilisticpredictionsgeneratedby
the MLA models provide a measure of the firm's financial position on a
continuous scale.
Earlier researchers(e.g., Ohlson [1980]) typically convert their prob-
abilistic MLA predictions into deterministic ones before they evaluate

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions
138 AMY HING-LING LAU

the predictions' accuracy. Deviating from this practice, this study uses
the probabilistic MLA predictions and a ranked probability scoring rule
to evaluate accuracy.
REFERENCES
ALTMAN, E. T. "Financial Ratios, Discriminant Analysis and the Prediction of Corporate
Bankruptcy." Journal of Finance (September 1968): 589-609.
, R. HALDEMAN, AND P. NARAYANAN. "ZETA Analysis: A New Model to Identify
Bankruptcy Risk of Corporations." Journal of Banking and Finance (June 1977): 29-54.
BEAVER, W. H. "Financial Ratios as Predictors of Failure." Journal of Accounting Research
(Supplement 1966): 71-102.
BERKMAN, J., D. BROWNSTONE, AND ASSOCIATES. Quail 4.0 Users' Manual. Berkeley:
Department of Economics, University of California, 1979.
BLUM, M. P. "Failing Company Discriminant Analysis." Journal of Accounting Research
(Spring 1974): 1-25.
DEAKIN, E. B. "A Discriminant Analysis of Predictors of Failure." Journal of Accounting
Research (Spring 1972): 167-79.
DIELMAN, T. E., AND H. R. OPPENHEIMER. "An Examination of Investor Behavior During
Periods of Large Dividend Changes." Journal of Financial and Quantitative Analysis
(June 1984): 197-216.
DONALDSON, G. Strategy for Financial Mobility. Boston: Division of Reseach, Graduate
School of Business Administration, Harvard University, 1969.
EMERY, G. W., AND K. 0. COGGER. "The Measurements of Liquidity." Journal of Account-
ing Research (Autumn 1982): 290-303.
EPSTEIN, E. S. "A Scoring System for Probability Forecasts of Ranked Categories." Journal
of Applied Meterology 8 (1969): 985-87.
FINANCIAL ACCOUNTING STANDARDS BOARD. Reporting Funds Flow, Liquidity and Finan-
cial Flexibility: Discussion Memorandum. Stamford, Conn.: FASB 1980.
GENTRY, J. A., P. NEWBOLD, AND W. WHITFORD. "Classifying Bankrupt Firms with Funds
Flow Components." Journal of Accounting Research (Spring 1985): 146-60.
HEATH, L. C. Financial Reporting and the Evaluation of Solvency. Accounting Research
Monograph no. 3. New York: AICPA, 1978.
JONCKHEERE, A. R. "A Distribution-Free k-Sample Test Against Ordered Alternatives."
Biometrika 41 (1954): 133-43.
JoY, M. O., AND J. 0. TOLLEFSON. "On the Financial Applications of Discriminant
Analysis." Journal of Financial and Quantitative Analysis (December 1975): 723-39.
LAU, A. H. L. "On the Prediction of Firms in Financial Distress with an Evaluation of
Alternative Funds-Flow Concepts." Ph.D. dissertation, Washington University, St. Louis,
1982.
McFADDEN, D. "Conditional Logit Analysis of Qualitative Choice Behavior." In Frontiers
in Economics, edited by P. Zarembka. New York: Academic Press, 1973.
MURPHY, A. H. "A Note on the Ranked Probability Score." Journal of Applied Meterology
10 (1971): 155-56.
, AND R. L. WINKLER. "Scoring Rules in Probability Assessment and Evaluation."
Acta Psychologica 34 (1970): 273-86.
NERLOVE, M., AND S. J. PRESS. Univariate and Multivariate Log-Linear and Logistic
Models. Santa Monica, Calif.: Rand Corporation, December 1973.
OHLSON, J. A. "Financial Ratios and the Probabilistic Prediction of Bankruptcy." Journal
of Accounting Research (Spring 1980): 109-31.
PETTIT, R. R. "Dividend Announcement, Security Performance and Capital Market Effi-
ciency." Journal of Finance (December 1972): 993-1007.
ZAVGREN, C. V. "Assessing the Vulnerability to Failure of American Industrial Firms: A
Logistic Analysis." Journal of Business Finance and Accounting (Spring 1985): 19-44.
ZMIJEWSKI, M. E. "Methodological Issues Related to the Estimation of Financial Distress
Prediction Models." Journal of Accounting Research 22 (Supplement 1984): 59-82.

This content downloaded from 128.83.63.180 on Mon, 08 Jun 2015 17:52:34 UTC
All use subject to JSTOR Terms and Conditions