Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Zavgren 1985 PDF

Download as pdf or txt
Download as pdf or txt
You are on page 1of 27

Journal ofBusiness Finance &Accounting, 12(1), Spring 1985, 0306 686 X $2.

50

ASSESSING THE VULNERABILITY


T O FAILURE OF AMERICAN INDUSTRIAL FIRMS:
A LOGISTIC ANALYSIS

CHRISTINE
V. ZAVGREN'

PURPOSE

This study specifically:


(1) uses the logit technique to develop and test a new bankruptcy model
which enumerates the signs of financial ill health for a five year period
prior to failure, and;
(2) develops a methodology for evaluating the significance of probabilities of
financial risk.

PRINCIPAL FINDINGS

The models proved highly significant with reference to both the R2and likelihood
ratio tests in detecting ailing firms up to five years prior to their failure.
By applying the entropy concept to the probabilities from the models, a
methodology is developed which objectively measures the quantity of informa-
tion in these signals. The information available from these models is significant
even five years prior to failure, and it increases up to the year immediately prior
to failure. This indicates a decrease in uncertainty over the fate of these firms
over this lead time.
The importance of individual variables is realistically assessable in this study
for the first time. The variables included in the models are determined empiri-
cally so as not to omit any important financial attributes. The results obtained
highlight the dimensions of financial data to which researchers and practitioners
should devote attention should they wish to understand the impact of managerial
activities on the firm's financial risk.
The profitability measure proves insignificant in any year, while the turnover
ratios are significant for the long run. The significance of the liquidity measure
in eF.rlier years and the negative sign of its coefficient indicates that the most
successful firms devote their resources to productive capital rather than liquid
* The author is Assistant Professor of Management at the Krannert Graduate School 01
Management, Purdue University, West Lafayette, Indiana. (Paper received May 1983, revised
May 1984)

19
20 ZAVGREN

assets. The acid test ratio proved to be highly significant with a negative coeffi-
cient in the first three years prior to failure. This points to the ability to meet
maturing obligations as an important factor in avoiding bankruptcy. The ratio
of long term debt to invested capital was found to be highly significant in all
years because the most reliable indication of a firm’s health or lack of health is
its use of debt.
When evaluating the classification and prediction error rates for all five
years, total error rates compare well with previous studies. The models perform
especially well in assessing the critical earlier years prior to failure.

SUMMARY OF PRIOR EMPIRICAL RESEARCH

Beaver (1965, 1966) pioneered empirical research in business failure prediction


using a univariate model. Although the univariate approach has certain short-
comings, especially a lack of integration of the various ratios, Beaver’s model
achieved a moderate level of predictive accuracy. Later multivariate studies
usually employed discriminant analysis. Altman (1968) initiated the multivariate
approach and several modifications of his basic technique followed. Deakin
(1972) improved the accuracy over that of Altman’s model by using all fourteen
ratios which Beaver had identified as good predictors of bankruptcy and a
probabilistic classification rule rather than a critical cut-off point. Blum’s (1974)
model incorporated variables expressed in terms of change over time. Edmister’s
function should be more stable than those of his predecessors, since the highly
correlated variables have been eliminated. Validation of this is difficult,
however, because the sample size is quite limited and is restricted to small
businesses. Wilcox (1971, 1973) developed the variables used in the discriminant
function by. deriving a random walk with drift formulation of transition to the
failure state. Diamond (1976) refined the multivariate model by using realistic
prior probabilities and error cost estimation, and by using an optimal discriminant
plane technique with a Bayesian classifier.
The conclusions of most of these studies are difficult to assess because they
play loose with the assumptions of discriminant analysis. Discriminant analysis
requires that the variables are normally distributed and the populations have
equal variance-covariance matrices. Otherwise, a linear classification rule and
the test for distinction between groups, usually Hotelling’s T’, become inappro-
priate. Non-normality also affects this test statistic and Eisenbeis indicates the
predictive accuracy of the function may be violated (Eisenbeis, 1977,pp. 875 - 7).
With the exception of Diamond (1976) and Altman el al. (1977), none of these
authors test their data against these assumptions; confusion over the purposes
of discriminant analysis sometimes leads to inappropriate attempts to assess the
meaning of individual coefficients. The linear combination of variables unique
to a particular discriminant function renders the importance of individual
coefficients impossible to assess.
ASSESSING VULNERABILITY T O FAILURE OF US INDUSTRIAL FIRMS 21
All of the discriminant analysis models identify differing sets of discriminating
variables [as many as thirty-two (Chen and Shimerda, 1980)l and most
models are specific to the samples from which they are developed. It would be
desirable therefore to have a theory which indicates the important dimensions
of financial information to include in these models; however none is available
besides the well-known theories of optimal capital structure.

STATISTICAL METHODOLOGY, SAMPLE AND VARIABLE SELECTION

Conditional Probability Models


Generally, conditional probability models estimate the probability of occurrence
of a choice or outcome; they depend on the attribute vector of the individual
and the choice or outcome set available. Although developed by a biologist
(Finney, 1952), they can assay the probability of commercial failure. See
Ohlson (1980), White and Turnbull’s unpublished analyses of the probability
of failure of railroad firms (1975), and industrial firms (1975), Santomero and
Vinso’s (1977) analysis of the failure of commercial banks, Chesser’s (1974)
study of commercial loan noncompliance, and Martin’s (1977) study of the
probability of bank failure.
Conditional probability models derive the probability of a dichotomous (or
polytomous) dependent variable by using coefficients on the independent
variables. These coefficients can be interpreted as the effect of a unit change in
an independent variable on the probability of the dependent variable. A
cumulative probability distribution is necessary to constrain the predicted
values within the acceptable [0,1] limiting values of probability distributions. If
the logistic function is used this constitutes the logit model. The logit model
uses the Sech2distribution (Maddala, 1977, pp. 163-4):

This distribution’s cumulative density is the logistic:


e”
F(u) = -
1 + ey

the logit model, estimated using maximum likelihood methods, is:

ComparativeAdvantages of Conditional Probability Models Over Discriminant Analysis


Deciding between discriminant analysis or a conditional probability model
depends largely on the intended use. Dichotomous classification(failing/healthy),
22 ZAVGREN

requires only discriminant analysis. But this dichotomous partition of the


outcome space is much less useful for an investor in capital stock, purchaser of
bonds, or a banker making a commercial loan decision than a cardinal evalua-
tion of financial risk. For many decisions, “the user may be capable of varying
levels of response to risk of failure,” (Martin, p. 269). For example, when
discussing commercial loan noncompliance Chesser (1975, p. 38) observes that
“noncompliance does not mean that a borrower will completely default on his
loan, but rather that some ‘work-out’ agreement will have to be arranged
which will result in settlement of the loan under conditions less tavorable to
the lender than those specified in the original agreement.” Assessment of the
likelihood of such an event makes possible such differential adjustments as
risk-premiums on interest rates and loan indentures.
Discriminant analysis also can generate a probability, but this usually involves
subjective assessment of the probability associated with a particular discrimi-
nant score. Martin (p. 258) employed a variant of discriminant analysis which
uses a maximum likelihood estimation technique to assess probabilities. Using
a logit model he tested the resqlfs of this estimation against the null hypothesis
that the probability of failure is equal to the prior probability in the population.
Martin found that both linear and quadratic discriminant functions had likeli-
hood functions significantly lower than the null hypothesis. This would mean
that the null hypothesis would provide a better probability estimate than either
discriminant function. The logit model, on the other hand, had a likelihood
function significantly higher than the null hypothesis, which indicates the logit
model provided significantly better probability estimates from the same data
(Martin, p. 266).
Martin provides additional evidence that the probabilities obtained from the
discriminant function may be very inaccurate, even though classification
accuracy is high. When a population contains strongly assymetric proportions
of groups, as with bankruptcy, the classificatory results improve by exaggerating
the size of the smaller group, even though by the maximum likelihood criterion
the discriminant function would be rejected. Thus, it is not surprising that the
use of a non-representative group biases the results of discriminant analysis
(Martin, p. 262), since most discriminant analysis studies use equal-sized
matched samples.
The coefficients derived from the conditional probability models estimate
representative effects of population parameters on the outcomes in the popula-
tion. If these coefficients are applied to the attribute vector of an individual firm
in the sample, the resulting index measures the “propensity to fail” (Martin,
1977, p. 257), the “vulnerability” to failure (Korobow and Stuhr, 1975, pp.
157-65), conditional on the firm’s attribute vector.
Many unobservable factors influence the vulnerability of an individual firm.
These include the unmeasured qualities of assets, the creative ability of manage-
ment, random events and the decisions of regulators and courts of law. Any
econometric model containing only financial statement information will not
ASSESSING VULNERABILITY ‘1’0 FAILURE OF US INDUSIRIAL FIRMS 23
predict with certainty the failure or nonfailure of a firm. As Martin recognized,
when discussing bank failure (p. 257): “These excluded variables (most of
which cannot be directly observed) determine how vulnerable, in terms of the
included variables, a bank would have to be in order to fail.” These factors
determine the “tolerance for vulnerability”, beyond which the firm will fail.
McFadden also discusses this issue in terms of “representative” characteristics
of the population; thus the prediction of an outcome for an individual will be
correct only if the representative element of his outcome function dominates
the idiosyncratic element (McFadden, p. 108).
Ohlson’s (1980) study uses a logit model and most closely precedes this one.
However, this study attempts to resolve several issues he does not adequately
address. First, Ohlson suffers from the lack of theoretical determination of his
model. He uses a conditional logit model to classify failing and healthy firms.
But he selects the independent variables without benefit of theory, assuring
problems similar to those observed for discriminant analysis. For example,
asset size is a variable with high significance in his model, but is also a scale
factor in other ratios. This renders independent conclusions about asset size
impossible to assess (Lev and Sunder, pp. 190- 3). He also fails to use a matched
sample technique, which would have controlled for inexplicit factors. In fact,
Ohlson calculated his error rate for the first year prior to failure using the same
sample he used to develop his model. Even so he found these error rates
disappointing since they are higher than those reported in previous studies.
These results are ambiguous. Finally, it is impossible to assess their dependence
on failure to capture the comprehensive dimensions of information about the
firm.
This study ameliorates this issue by determining the attribute vector empiri-
cally through factor analysis. Thus, no significant attribute of these firms
should be omitted. Further extensions of the methodology are to analyze the
pattern of importance of these variables, and also the significance of the
probabilities generated by the models.

Selection of Variables and Sample


Two recent studies (Pinches, Mingo and Carothers, 1973; and Pinches, Eubank,
Mingo and Carothers, 1975) apply factor analysis. They identify the ratios
which constitute the principal independent dimensions of financial statement
data, and assess the stability of these dimensions over the short-run and the
long-run. Both studies identify the same sets of orthogonal factors, selected
from a similar, large group of ratios. The authors found them relatively stable
over both the short run and the long run, indicating that they are stable dimen-
sions of financial information. Each factor contained at least one ratio with a
very high factor loading (correlation with the factor). This indicates that most
of the information in that ratio is related to the factor.
This study employs all the ratios which Pinches et al. identified as most strongly
related to their seven factors, with the exception of the current ratio, which is a
24 ZAVCREN

measure of short-term liquidity. The current ratio increases in proportion as


the failing firm’s unsaleable inventories accumulate, and so it provides a
misleading measure of liquidity. If, however, one substitutes the current ratio
with the acid test ratio, one can then ignore inventories, and derive a clearer
picture.
The variables chosen represent an empirically determined attribute vector of
the firm, according to which one can make unprejudiced, accurate distinctions
between failing and healthy firms. Table 1 indicates the variables used here,
along with the factors to which they correspond.

Table 1
Financial Ratio Data Set

Factor
Ratio Factor Loading
(1) Total Income/Total Return on Investment 0.97
Capital
(2) Sales/Net Plant Capital Turnover 0.95
(3) Invent ory/Sales Inventory Turnover 0.97
(4) Debt/Total Capital Financial Leverage 0.99
(5) Receivables/Inventory Receivables Turnover - 0.99
(6) Quick Assets/Current Short-term Liquidity 0.81
Liabilities
(7) CashITotal Assets Cash Position 0.91

Certain expectations can be developed a p h n ’ over the sign on the coefficients


attributed to these variables, and the pattern of the strength of their significance
with respect to failure over the five year study period. Profitability provides a
measure of return on investment, and can be expected to be a significant
measure of a firm’s health. The literature of financial economics points out the
importance of profitability (as well as capital structure and liquidity) as variables
(see, for example Tinsley, 1970; or Stiglitz, 1972). Failing firms are expected to
have higher financial leverage than healthy ones, since inability to meet high
fixed debt service obligations is frequently a precipitating factor in a firm’s
demise. The literature of optimal capital structure highlights the use of debt as
increasing the riskiness of the firm (see Modigliani and Miller, 1958; Tinsley,
1970; Smith, 1972; Stiglitz, 1972; and Wilcox, 1971).
Efficiency ratios would be expected to have the greatest significance in the
long run because they measure the ability of the firm to use assets to full capacity.
Although “turning assets over” improves long run profitability, it may be
ASSESSING VULNERABIIXIY 1’0FAILURE O F U S INDUSTRIAL FIRMS 25
difficult to effect and distinguish changes in these ratios over the short term,
especially in the case of fixed asset turnover. This should hold for inventory and
receivables turnover as well, where efficiency also results from long-run deci-
sions. Similarly, capital investment and marketing policy, which would affect
these ratios, are not amenable to rapid short-run changes; and the effect of such
policies on operations will be observed mostly over the long run. Thus, these
variables should be significant in the financial failure model also only over this
horizon. If one focuses on the significance of the numerator of the capital turn-
over ratio (increasing sales for a given plant base) then the coefficient would be
expected to be negative. Even if a firm neglects to invest in capital assets, it
would yield an impressive capital turnover over the short-run. Then, after a
period of time, these firms will become less efficient than firms more vigorous
in investing, and so are more likely to fail. Receivables turnover and inventory
turnover would both be expected to have positive signs for their coefficients;
clearly when receivables increase faster than inventories or inventories faster
than sales, failure looms.
The acid test ratio indicates the firm’s ability to meet currently maturing
obligations. This would be expected to be particularly important over the short
run, as an inadequate reserve of quick assets could precipitate bankruptcy.
Current assets to total assets measures the relative liquidity of the firm’s
assets. This ratio would be expected to be critical immediately prior to failure
and have a negative coefficient, since only liquid assets can generate cash to
cover obligations. However, over the long-run, successful firms try to
minimize liquidity (subject to transactions needs) in order to channel funds into
productive assets. Overcommitment of funds to investment or inadequate
anticipation of upcoming needs for liquidity may easily precipitate crisis.
The firms included in the sample were selected from the firms on the
COMPUSTAT New York Stock Exchange, Over-The Counter, and Research
tapes. The failed firms were identified as having filed Chapter X or Chapter
XI bankruptcy proceedings in the period of 1972 - 78 from the F and S Index of
Corporate Changes. Of the firms on the COMPUSTAT tapes, 130 could be iden-
tified as having failed by this definition during this period. Of these 130 firms,
69 had the manufacturing industry classifications (SIC codes 2200 - 3940), of
which 16 were dropped because of missing data.
The failed firms were matched with healthy firms by four digit industry code
and total asset size as of the beginning of the sample period. Smaller firms are
generally presumed to have failure rates greater than larger firms, a presump-
tion Ohlson’s study supports, so a universally valid measure requires controll-
ing for firm size. Three firms could not be controlled, and for another five data
was out of date. This left a sample of 45 failed and 45 healthy firms. Table 2
contains the names of the firms and their year of failure. The sample of failed
firms is not random. The 45 failed firms constitute the entire population of failed
firms in the period 1972 to 1978 for which data is available. Thus cross sectional
generalization of the results is not an empirical issue. Generalization over time
Table 2
Sample of Failed and Matching Firms

Failed Finns Matching Finns


Ind. Co. Compav Bankruptcy Last Id. Co. Company Indurtry
No. No. name Petition Data No. No. Name Name
Dcuc
2200. 266057 Duplan 8 76 1975 2200. 12347 Albany Intl. Corp. Textile Mill Products
2200. 525030 Lehigh Valley Railrd 7 77 1976 2200. 88665 Bibb Co. Textile Mill Products
2270. 817460 Sequoyah Industries 11 74 1972 2270. 574803 Masland (O.H.) & Sons Floor Covering Mills
2300. 101241 Botany Industries 4 73 1971 2300. 95293 Blue Bell Inc. Apparel and Other Finished Pds.
2300. 269461 Eagle Clothes 3 78 1976 2300. 471016 Jantzen Inc. Apparel and Other Finished Pds.
2450. 959078 Western Orbis 6 76 1974 2450. 731588 Poloron Products Inc. Wood Buildings - Mobile Homes
2450. 202651 commodore 11 76 1976 2450. 206039 Conchemco Inc. Wood Buildings - Mobile Homes
2650. 703181 Paterson Parchment 4 77 1973 2650. 828658 Simkins Industries Paperboard Containers - Boxes
3069. 763121 Richardson Company 4 75 1974 3069. 24591 American Biltrite Fabricated Rubber Prods. NEC
3140. 358604 Frier Industries 7 78 1975 3140. 864261 Suave Shoe Corp. Footwear Except Rubber
3140. 26367 American Girl Fashion 2 75 1973 3140. 498552 Kleinens Inc. Footwear Except Rubber
3250. 233045 DCA Development 2 73 1971 3250. 826750 Sikes Corp-C1 A Structural Clay Products
3270. 577438 Maule Industries 5 77 1975 3270. 882491 Texas Industries Inc. Concmte Gypsum & Plaster
3310. 11608 Alan Wood Steel 1 78 1976 3310. 549866 Lukens Steel Co. Blast Furnaces & Steel Works
3310. 880447 Tennessee Foreign St 12 77 1976 3310. 32519 Anadite Inc. Blast Furnaces & Steel Works
3341. 829880 Sitkin Smelting 3 78 1977 3341. 743396 Proler International Corp. Second Smelt.-Ref. In Nonfer Mt.
3350. 755400 Reading (Ind.) 5 74 1973 3350. 262093 Driver Harris Co. Rolling & Draw Nonfer Metal
3350. 903200 UOP 10 75 1974 3350. 761406 Revere Copper & Brass Inc. Rolling & Draw Nonfer Metal
3449. 389370 Gray Manufacturing 10 75 1974 3449. 458884 International Aluminium Misc. Metal Work
3480. 749482 REDM 12 72 1972 3480. 754713 Raymond Industries Ordinance & Accessories
3499. 36159 Anodyne 10 76 1972 3499. 942662 Watsco Inc. Fabricated Metal Prods. NEC
3500. 9124 Air King 6 75 1973 3560. 7680 Aero-Flow Dynamics Inc. General Indust. Mach. & Equip.
3560. 268039 Dynamics Corporation 12 74 1974 3560. 487314 Keene Corp. General Indust. Mach. & Equip.
Failed F i m -Matching F i m
Id. CO. COmpanY BakjJtCY Lust Id. CO. Company Zndurtry
No. NO name Petition Dala No. No. Name NomC
Dalc
3570. 738102 Potter Instrument 1 78 1974 3570. 130163 California Computer Prod. Office Computing & Acctq. Mach.
3570. 250649 Detect0 Scales 7 76 1973 3570. 933696 Wang Laboratories Office Computing & Acctq. Mach.
3610. 923601 Vogue Instruments 5 75 1972 3610. 4644 Acme Electric Corp. Elec. Transmission & Distr. Eq.
3651. 460550 International Video 5 77 1977 3651. 296677 Esquire Radio & Electron Inc. Radio-TV Receiving Sets
3662. 372298 Genisco Technology 12 78 1972 3662. 808655 Scientific-Atlanta Inc. Radio-TV Transmitting Equip.-Ap.
3662. 376424 Cladding 12 78 1976 3662. 313855 Federal Signal Corp. Radio-TV Transmitting Equip.-Ap.
3662. 448519 Hy-Gain Electronics 11 78 1977 3662. 477178 Jetronic Industries Inc. Radio-TV Transmitting Equip.-Ap.
3662. 286155 Electrospace 4 74 1972 3662. 387100 Granger Associates Radio-TV Transmitting Equip.-Ap.
3662. 103025 Bowmar Instrument 4 77 1976 3662. 478358 Johnson (E.F.) Co. Radio-TV Transmitting Equip.-Ap.
3679. 740274 Precision Instrument 3 75 1972 3679. 296650 Espey Mfg. &ElectronicsCorp. Electronic Components NEC
3679. 683816 Cotel 1 77 1976 3679. 205826 Comtech Telecommunication
Corp. Electronic Components NEC
3679. 417434 Harvard Industries 8 76 1971 3679. 812098 Sealectro Corp. Electronic Components NEC
3714. 858552 Stellar Industries 4 75 1973 3714. 977487 wix Corp. Motor Vehicle Parts-Accessories
3714. 302419 FDI 12 78 1976 3714. 866713 Sun Electric Corp. Motor Vehicle Parts-Accessories
3714. 521687 Leader International 12 73 1972 3714. 443073 Howell Industries Inc. Motor Vehicle Parts-Accessories
3728. 7752 Aerodex 9 76 1974 3728. 422452 Heath Tecna Corp. Aircraft Parts & Aux. Equip.
3792. 683714 Open Road Inaustries 11 76 1973 3792. 295360 Vesely Co. Trans1 Trailers & Campers
3811. 171142 Chronetics 3 76 1972 3811. 553777 MTS Systems Corp. Engr. Lab & Research Equip.
3861. 703827 Pavelle 5 73 1972 3861. 35310 Anken Industries Photographic Equip. & Suppl.
3940. 364847 Garcia 8 78 1977 3940. 451650 Ideal Toy Corp. Toys & Amusement Sport Goods
3940. 858518 Stelber Industries 3 76 1974 3940. 761355 Revell Inc. Toys & Amusement Sport Goods
3940. 602805 Miner Industries 3 78 1975 3940. 743385 Progroup Inc. Toys & Amusement Sport G o o d s
28 ZAVGREN

is desirable however, and this will be assessed below with two years of data from
1979 and 1980.
The failure date reflects either the date of bankruptcy filing or the last period
for which financial data was published, whichever was earlier. Assuming that a
firm soon to file bankruptcy ceases publishing financial statements, the actual
petition filing is usually a foregone conclusion.

ESTIMATION RESULTS

Table 3 presents the estimation results for the logit models for one to five years
prior to the failure date. (Probit models based on the normal distribution pro-
duce similar results. This would be expected as the logit distribution differs
from the normal distribution only in being slightly more platycurdic; the results
are not presented here.) A discussion follows on the significance of the
estimated models; the significance of the probabilities; the significance of
individual coefficients; and the classification and prediction of error rates.

Significance of the Models


R2 measures the proportion of variation in the observed dichotomous variable
which the predicted index explains. Since R2is a measure of fit of a linear rela-
tion, a high R2 is not expected for the non-linear probit and logit estimations.
Only if the predicted values were all 0 or 1 could R2approximate 1 (Pindyck
and Rubinfeld, 1976, p.225). Morrison shows that the upper bound for R2 for
binary probabilistic models is 113, (1972, pp.68-71) where the true pro-
babilities are uniformly distributed across the [0.1] interval. The R' values
reported in Table 3 appear to be quite high for years one and two. But the R2's
for the remaining years seem acceptable, especially considering that the data is
cross-sectional and reflects only financial statement data.
The likelihood ratio test (Maddala, p. 179-80) provides a more definitive
test of the strength of the probit or logit model. It involves the null hypothesis
that the entire model is insignificant, or that all the coefficients are
insignificantly different from zero. Let L(0') be the value of the likelihood
function estimates for the original model (unrestricted). Let L(B) be the value of
the likelihood function with the restriction imposed that all coefficients are
zero. The likelihood ratio is then defined as:

Define the test statistic as - 2 log A which is asymptotically distributed as Chi-


square with degrees of freedom equal to the number of independent restrictions
imposed. The likelihood ratio test statistic for the probit and logit models
Table 3
Logit Estimation Results: Years 1 - 5

YU
Awrto Inn. Rec. Cash/ Acid Rctwn Debt Asset Likelihood m
X
Failure VmiOblcName Intercept Twn. Turn. Ass& Tat CapW Caflirol Turn Rz Raiio
1 Coefficient - 0.23883 0.00108 0.01583 0.10780 - 0.03074 - 0.00486 0.04350 - 0.00110 0.4479 45.7109
Asymptotic t - 0.1709 0.0482 1.6587 1.9328 - 3.8095 - 0.8809 3.3281 - 0.6762
Sig. level 7% 2% 90% 95% 99% 62% 99% 50% 99%
2 Coefficient -2.61060 0.04185 0.02215 0.11231 -0.02690 -0.01440 0.04464 0.00063 0.4529 46.9320
Asymptotic t - 1.5850 1.4371 1.8682 1.5924 -3.0651 -0.5740 3.3401 0.2560
Sig. level 89% 85% 94% 89% 99% 43% 99% 21 % 99%
3 Coefficient - 1.51150 0.06257 0.00829 0.4248 -0.01549 0.00519 0.01822 O.ooOo2 0.2665 25.8299
Symptotic t -1.0666 2.0364 1.2754 1.0345 -2.5817 1.2902 1.6446 0.01521 C
T
Sig. level 71% 96% 80% 69% 99% 80% 90% 0% 99%
E
4 Coefficient - 5.9457 0.09157 0.01667 0.05917 -0.00410 0.01950 0.04100 0.00363 0.2391 22.4438
Asymptotic t - 3.1765 2.9271 1.9837 1.1397 0.7759 1.6589 2.7322 1.6706
Sig. level 99% 99% 95% 75% 56% 90% 99% 90% 99%
5 Coefficient - 6.8766 0.08835 0.00692 0.15786 0.00018 - 0.02301 0.04371 0.00798 0.3207 30.2534
Asymptotic t -3.5209 2.6453 0.85248 2.1933 0.0599 - 1.91566 2.5403 2.7270
Sig. level 99% 99% 60% 97% 4% 64% 99% 99% 99%
30 ZAVGREN

reported in Table 3 all indicate significance at the 0.995 confidence level when
compared with the Chi-square distribution with seven degrees of freedom.
The important issue here is whether the models clearly distinguish between
failing and healthy firms. The likelihood ratio test is a strong indicator of this,
as are Figures 1 -5. The probabilities of the two samples, are shown to diverge
significantly by their respective bar graphs. The failing (healthy) sample is
clearly skewed toward the higher (lower) probabilities of failure. In years one,
two, and live, the modal probabilities for failing firms are between 0.90 and
1.OO.In year three the modal probability is between 0.70 and 0.80, and in year
four it is between 0.60 and 0.70. For the healthy firms, the modal probability of
failure was less than 0.10 for years one and two. Modal values for years three
and five were 0.20 to 0.30, and lor year four were 0.30 to 0.40.

Shannon's Entropy Theory Applied to Probabilistic Financial Failure Models


Shannon (1948) developed information theory for analyzing the transmission
of information through a communication network or channel. Theil (1967)
applied it to economics; Lev (1969) to accounting. This section applies infor-
mation theory to measure the information content of the predictions from the
model. It heightens the usefulness of the probabilities. Qualitative responses
are possible once a cardinal measure of risk is available. Such responses can be
made with more confidence if some objective measure is available of the
significance of probability changes over time and of cross sectional differences
in probabilities between firms.
Seen in the information theory context, the probability estimates generated
by the logit models are messages from an information system. The quantity of
information in each message can be measured by its ability to reduce uncer-
tainty. 'Entropy' is defined as the degree of uncertainty over the occurrence of
an event. This concept originates in statistical thermodynamics, and its mean-
ing has been subject to several interpretations (Georgescu-Roegen, 1971 ,
Cp.6), but the entropy concept applied here will be that developed by Shannon
(1 948, pp.379-423, 623 -656).
Shannon developed his concept out of an interest in determining the quantity
of information supplied by a probabilistic experiment. Thus he sought a func-
tion which when applied to the probabilities from an experiment would yield
the most reasonable measure of the quantity of information supplied by the
experiment. Such a function would be required to be additive; permit calcula-
tions of conditional measures similar to those appropriate for probabilities; and
be a decreasing function of the probability of an event. The function which best
fits these requirements is a logarithm of the probabilities; it results in the expec-
tation:

"
H" = H"(p,,. . . P") = - 2 p t log p t .
t- I
(5)
Figure 1
Estimated Probability of Failure for One Year Prior to Failure Date - Failed Firms and Non-Failed Matched Sample

Frequency of
Occurrence
Out of Sample 12
of 45 Failed 11
and 45 Non-Failed
Firms 10
9

8
7
6
5

3
2

= failed
- I
a o<p<o.10 0.20 <p <0.30 0.40<p <0.50 0.60 <p <0.70 0.80<p ~ 0 . 9 0
0.10 <p<0.20 0.30 <p<O.40 0.50 <p<O.SO 0.70 <p <0.80 0.90<p< 1.o
0 = non-failed
Probability of Failure
Figure 2 W
lu
Estimated Probability of Failure for Two Years Prior to Failure Date - Failed Firms and Non-Failed Matched Sample

Frequency of 15
Occurrence 14
Our of Sample
of 45 Failed 13
and 45 Non-Failed
Firms 12
11
10
9
N
8 >
<
n
7 7J
m
6
z
5

2
1

= railed
- 0.40< p < 0 . 5 0 0.60<p<0.70 0.80 <p<0.90 ~-
I 0
I
0 < p <0.10 0.20 c p <0.30
O.lO<p<0.20 0.30<p<0.40
d 0.50<p<0.60 0.70<p<0.80 0.90<p< 1 .O
= non-failed
Probability of Failure
Figure 3
Estimated Probability of Failure for Three Years Prior to Failure Date for Failed Firms and Non-Failed Matched Sample
$
(I:
Frequency of 15
Occurrence
Out of Sample 14
of 45 Failed
and 45 Non-Failed l3
Firms 12
11

10
9
8
7

5
4

3
n
L

1
Figure 4 w
rp

Estimated Probability of Failure for Four Years Prior to Failure Date for Failed Firms and Non-Failed Matched Sample
15
Frcquency of 14
Occurrence
Out of Sample 13
of 45 Failed
and 45 Non-Failed 12
Finns
11

10

8
7

3
2
1

-
= failed 0 O<p<O.lO 0.20<p<0.30 0.40<p<0.50 0.60 <p<0.70 0.80 <p<O.W
O.lO<p<0.20 0.30<p<0.40 0.50<p<0.60 0.70<p<0.80 O.W<p< 1.0
0 = non-failed
Probability of Failure
1
ASSESSING VULNERABILI'I'Y '1'0 FAILURE OF U S INDUS'I'RIAL FIRMS 35

B
5

-
B
.-
m
LL
t:
z"
T
C
m
m
E
.-
LL

0 -.
I mv"
V 0

55
m
m
;
6J
2 2
0 0
> V
iz
2
L mv" 0
36 ZAVGREN

This expectation is referred to as Shannon's entropy (Guiasu, 1977, pp. 1 - 15),


a value-free indication of the number of messages carried by an information
channel.
Shannon's entropy will be employed as a n exposf measure of the amount of
uncertainty over the occurrence of a n event before and after the receipt of a
message. A greater entropy in the probability distribution of an event will be
taken to imply a greater degree of surprise if that event were to occur. Thus,
entropy will be a decreasing function of the probability of an event. T h e units of
the entropy measure resulting from using natural logs are calls nits. Let
entropy be represented as h, thus:

he) = In (VP) (6)


If the outcome as regards a binary event s is unknown, the entropy cor-
responding to each state would be h e ) and h(1 -p), respectively. Each entropy
can be weighted by its probability of occurence, to give the expected entropy,
or (Theil, p.639):

The expected entropy gives an ex ante estimate of the uncertainty remaining


over the occurrence of an event. With the true outcome known, the entropy of
the message becomes an expost measure (in the same sense that prediction error
rates are on exposf measure).
The models used here provide sequential estimates of the probabilities of
failure as subsequent messages come in. T h e positive or negative increment in
information after the receipt of the second message becomes:
In lip2 - In l/pI = lnp,/p, (9)
Even if the state of nature remains obscure after the receipt of the last
message, the expected information can be calculated. Given two states and two
consecutive messages, the expected information is (Theil, pp.641-642):

Therefore, given that we know which firms have failed, how much informa-
tion is measured in the predicted probabilities? For the fifth year prior to
failure, the remaining uncertainty over the probabilities of failure is given by:

As succeeding predictions change, the information in these predictions is


similarly determined. The decrease (increase) in entropy after the receipt of the
succrt:ding messages is simply calculated according to equation (9). ?'able 4
presents the changes in entropy over tht: live- year period.
ASSESSING VULNERABILITY T O FAILURE OF US INDUSTRIAL FIRMS 37
Table 4
Information Content in Nits of the Logit Model, Failed Firms Years 1 - 5

4 1.56 1.55 0.01 - 0.21 0.22 - 0.51 0.73 0.04 0.69


9 0.00 0.00 0.00 -0.16 0.16 0.07 0.09 0.06 0.03
16 3.51 1.10 2.41 0.20 2.21 0.65 1.56 -0.27 1.83
22 0.65 - 0.37 1.02 0.05 0.97 0.08 0.89 -0.08 0.97
24 1.56 1.23 0.33 0.25 0.08 - 0.09 0.17 0.05 0.13
34 0.05 0.05 0.00 - 1.20 1.20 0.57 0.63 0.61 0.02
37 0.87 0.87 0.00 - 0.53 0.53 0.00 0.53 -0.27 0.80
47 0.24 0.68 0.92 0.50 0.42 - 0.22 0.63 -1.49 2.12
49 0.00 - 0.67 0.67 - 0.13 0.80 0.76 0.04 0.03 0.01
56 0.29 - 0.07 0.36 - 0.09 0.45 0.32 0.13 -0.02 0.15
57 0.20 - 0.05 0.25 - 0.04 0.29 - 0.19 0.48 -0.37 0.84
58 0.36 0.11 0.25 - 0.05 0.30 - 0.03 0.33 0.04 0.29
59 0.07 0.24 0.31 - 0.63 0.94 - 0.44 1.39 -0.45 1.83
60 0.15 - 0.48 0.63 0.04 0.60 0.07 0.53 -0.11 0.63
61 0.21 0.00 0.21 0.06 0.15 - 0.07 0.22 -0.15 0.37
62 0.45 0.26 0.19 - 0.66 0.84 0.46 0.39 -0.53 0.92
63 0.21 - 0.05 0.26 - 0.71 0.97 0.47 0.49 -0.26 0.76
64 0.97 0.55 0.42 0.03 0.39 - 0.53 0.92 0.71 0.21
65 0.22 - 0.01 0.24 0.50 0.73 0.02 0.71 -1.69 2.41
66 0.39 - 0.39 0.78 -0.17 0.94 0.40 0.54 0.19 0.36
67 0.26 - 0.45 0.71 - 0.60 1.31 0.92 0.39 0.20 0.19
68 0.05 0.01 0.04 - 0.02 0.06 - 0.04 0.11 -0.11 0.21
69 0.03 0.02 0.01 - 0.01 0.02 0.02 0.00 0.00 0.00
70 0.58 - 0.73 1.31 0.55 0.76 - 0.52 1.27 0.33 0.94
71 0.30 -0.16 0.46 0.19 0.27 - 0.31 0.58 0.22 0.36
72 0.21 0.15 0.06 0.02 0.04 -0.15 0.19 0.12 0.06
73 0.29 0.05 0.24 -0.11 0.34 - 0.22 0.56 -0.02 0.58
74 0.01 -0.11 0.12 - 0.24 0.36 - 0.38 0.73 0.33 0.40
75 0.63 0.28 0.36 - 0.72 1.08 - 0.35 1.43 0.71 0.71
76 0.27 0.16 0.12 0.10 0.02 0.00 0.02 0.02 0.00
77 0.04 - 0.02 0.06 - 0.59 0.65 0.38 0.27 0.15 0.13
78 0.62 0.53 0.08 - 0.35 0.43 0.03 0.40 0.00 0.40
79 0.97 0.44 0.53 - 0.09 0.62 0.00 0.62 0.00 0.62
80 0.06 - 0.32 0.39 0.06 0.33 - 0.18 0.51 0.10 0.42
81 0.13 -0.11 0.24 - 0.07 0.30 - 0.26 0.56 -0.04 0.60
82 0.04 - 0.05 0.09 - 0.22 0.31 - 0.58 0.89 0.60 0.29
83 0.39 - 2.02 2.41 1.96 0.45 0.18 0.26 -0.04 0.22
84 0.11 0.04 0.06 0.00 0.06 - 0.52 0.56 -0.73 1.31
85 0.13 - 0.15 0.27 0.01 0.26 - 0.03 0.29 -0.07 0.36
86 1.61 - 0.60 2.21 0.61 2.81 0.69 2.12 1.47 0.65
87 0.01 0.00 0.01 0.01 0.02 - 0.23 0.25 0.05 0.20
88 0.04 -0.16 0.20 - 0.36 0.56 - 1.21 1.77 1.47 0.30
89 0.54 0.48 0.06 -0.11 0.17 - 0.09 0.26 0.16 0.11
90 0.56 0.19 0.37 - 0.50 0.87 0.68 0.19 0.14 0.05
Total 20.05 19.82 25.55 25.85 24.56
Average
Entropy 0.46 0.02 0.44 -0.13 0.57 0.00 0.57 0.01 0.56
(nits)
38 ZAVGREN

The total decrease in entropy over the five year period is the difference bet-
ween entropy values for the fifth and first years. For the failed firms, the
average entropy corresponding to the probability predictions for the five year
period decreases by the amount of 0.10 nits. This corresponds to the decrease
in uncertainty over the five year period, or the differential quantity of informa-
tion due to the posterior signals. T h e decrease in entropy over the four year
interval is 0.11 nits. This indicates that for the failed firms, 18 per cent more
information than that already available from the model for the fifth year prior
to failure was picked up during the five year lead time.
The entropy corresponding to the prior probability (year 5) must be inter-
preted in a relative sense. In order to compare alternative information systems,
the quantity of uncertainty remaining after the receipt of this measure must be
compared to the quantity of uncertainty remaining after the receipt of a signal
from an alternative information system. Since no such estimates are available,
comparison cannot currently be made. T h e level of magnitude of the informa-
tion change measure, however, might be compared to Lev’s ‘balance sheet
decomposition information measure’ (Lev, 1971, pp. 106- 107). T h e
magnitude of these measures is roughly similar. It must be remembered,
however, that Lev’s results are not derived from true probabilities, but from
changes in relative proportions of certain balance sheet items. As such, Lev’s
model cannot distinguish between changes in composition due to successful
growth and changes in composition due to impending failure.
Looking at all this positively, could healthy firms use these models to
measure, even advertise their health? For a five year prediction the remaining
uncertainty over the probability of health is given by the entropy which cor-
responds to the predicted probability of health 1 - p:
h(1 -p) = ln(1/1 -p) (12)
Similarly, a measure of entropy was calculated for each succeeding year. T h e
total decrease in entropy over the five year lead period, reported in Table 5, is
0.08 nits, on an average basis.
Over the four year lead time the measure is 0.14 nits. For the healthy firms,
the uncertainty of their fate shrinks by 16 per cent during the five year lead
time. T h e fate of the healthy firm is relatively sure over the five year test period,
and becomes an even better risk according to the information from these
models. The information content derived from this study greatly exceeds that
reported by Lev for his sample of healthy firms. T o corroborate these results,
several interesting properties of the logit model indicate that entropy is an
appropriate measure of its success.
The predicted index is restricted by the logistic distribution, In (PI1 - p). l h i s
index corresponds to the difference between the two entropies, uiz:

= In [pll -PI
ASSESSING VULNERABILITY TO FAILURE O F US INDUSTRIAL FIRMS 39
Table 5
Information Content in Nits of the Logit Model, Nonfailed Firms Years 1 - 5

Year Prior io Failure

I 2 3 4 5
Obsmation h(1 -p) Ah(1 - p ) h(l -p) Ah(1 -p) h(l -p) Ah(1 -p) h(l -p) Ah(1 -p) h(1 -p)
2 0.69 0.36 0.33 -0.07 0.40 - 0.18 0.58 0.32 0.26
3 0.37 0.14 0.24 0.01 0.22 - 0.30 0.53 -0.02 0.54
5 1.17 -0.22 1.39 0.39 0.99 - 0.97 1.97 0.54 1.43
6 1.47 0.99 0.48 0.25 0.22 - 0.77 0.99 0.40 0.60
7 0.37 0.11 0.26 -0.15 0.42 0.02 0.40 -0.18 0.58
8 0.07 0.03 0.04 -0.47 0.51 0.11 0.40 -0.03 0.37
10 0.22 0.14 0.08 -0.26 0.34 0.11 0.24 -0.02 0.26
11 0.19 0.17 0.07 -0.20 0.27 0.17 0.11 0.00 0.11
12 0.03 -0.01 0.04 -0.21 0.25 - 0.30 0.54 0.05 0.49
13 0.45 -0.12 0.56 -0.26 0.82 0.49 1.31 0.82 0.49
14 0.14 -0.05 0.19 -0.31 0.49 0.18 0.31 0.03 0.29
15 0.19 -0.29 0.48 0.06 0.42 0.04 0.37 0.14 0.24
17 0.03 -0.40 0.43 -0.17 0.60 - 0.34 0.94 0.67 0.27
18 0.76 0.18 0.58 0.15 0.43 - 0.39 0.82 0.49 0.33
19 1.43 0.67 0.76 0.08 0.67 0.04 0.63 0.23 0.40
20 2.81 0.51 2.30 1.61 0.69 0.48 0.21 0.00 0.21
21 0.25 -0.05 0.30 -0.35 0.65 0.31 0.34 -0.22 0.56
23 0.31 -0.03 0.34 -0.20 0.54 0.35 0.20 0.00 0.20
25 0.63 -0.11 0.53 0.02 0.51 - 0.24 0.76 0.10 0.65
26 0.94 -0.83 1.77 0.27 2.04 - 0.49 2.53 0.00 2.53
27 0.12 -0.05 0.16 -0.18 0.34 0.00 0.34 -0.29 0.63
28 0.03 -0.03 0.06 -0.02 0.08 - 0.02 0.11 -0.05 0.16
29 1.43 0.15 1.27 -0.50 1.77 1.12 0.65 0.44 0.21
30 0.37 -0.06 0.43 -0.05 0.48 - 0.69 1.17 0.72 0.45
31 0.07 0.03 0.04 -0.08 0.12 - 0.20 0.31 0.15 0.16
32 0.36 0.01 0.34 0.12 0.22 - 0.28 0.51 0.34 0.17
33 0.24 0.18 0.06 -0.28 0.34 0.07 0.27 -0.22 0.48
35 0.05 -0.03 0.08 0.06 0.14 - 0.07 0.21 -0.16 0.37
36 0.00 -0.03 0.03 -0.02 0.05 - 0.12 0.17 0.00 0.17
38 0.67 0.45 0.22 -0.80 1.02 0.53 0.49 0.18 0.31
39 0.08 -0.10 0.19 -0.13 0.31 - 0.46 0.78 0.35 0.43
40 0.00 -0.05 0.05 0.05 0.00 - 0.34 0.34 0.03 0.37
41 0.87 0.11 0.76 -0.04 0.80 - 0.12 0.92 0.19 1.11
42 0.43 0.39 0.04 -0.03 0.07 -0.10 0.17 -0.04 0.21
43 0.15 -0.30 0.45 -0.25 0.69 0.13 0.56 -0.35 0.92
44 0.36 - 1.68 2.04 0.99 1.05 0.59 0.46 0.21 0.25
45 0.30 -0.10 0.40 -0.38 0.78 0.14 0.63 0.11 0.53
46 0.37 0.04 0.33 -0.29 0.62 0.33 0.29 -0.08 0.37
48 0.14 0.08 0.06 -0.63 0.69 0.26 0.43 0.13 0.30
50 0.11 -0.20 0.30 -0.13 0.43 0.27 0.16 -0.22 0.39
51 0.12 -0.27 0.39 0.10 0.29 - 0.53 0.82 - 1.48 2.30
52 0.40 0.19 0.21 0.16 0.37 0.01 0.36 -0.75 1.11
53 0.04 0.02 0.02 -0.17 0.19 - 0.01 0.20 -0.02 0.22
54 0.05 0.04 0.01 -0.35 0.36 0.04 0.31 0.03 0.29
55 0.05 0.02 0.03 -0.59 0.62 - 0.10 0.71 0.49 0.22
Total 19.33 - 19.14 23.31 25.55 22.94
Average
Entropy 0.43 0.00 0.43 -0.08 0.52 - 0.05 0.57 0.06 0.51
(nits)
40 ZAVCREN

It is also significant that the first derivative of the expected quantity of informa-
tion is minus the logit index; this indicates that the logit prediction is the
negative of the rate of change in the quantity of information as the probability
of the event changes,:
dH/dp = d (p In 1/p (1 - p) In 1/( 1 - p))/dp
= - In pl(1 -p) (16)
Call the information content of the occurence of the states a loss function. It
evaluates the predictions generated by the model, in the same way that the
mean square error evaluates the predictions from linear regression models.
Using maximum likelihood estimation minimizes the average loss. Further,
the entropy of the distribution equals the expected loss, and is minimized by the
true probabilities (Brelsford and Jones, 1967, pp.570-576).

Sign;f;ance of Individual CoeJicients


The asymptotic t-statistic corresponding to each variable’s coefficient evaluates
the significance of the coefficients in each year’s model. This compares to the
values of the standard normal distribution. Table 3 indicates the significance of
coefficients from year to year. Long-term debt to invested capital is significant
at the 99 per cent confidence level for four years, and in year four it is signifi-
cant at the 89 per cent confidence level. Clearly, failing firms use leverage more
than healthy firms.
Efficiency ratios have the greatest significance in the long run because they
measure the ability of the firm to use assets to full capacity. Although ‘turning
assets over’ improves long run profitability, it is difficult to distinguish short
term changes, especially in the case of fixed asset turnover. T h e same is true for
inventory and receivables turnover. Although the Pinches, et al. analyses
indicate the importance of these ratios in accumulating overall financial data,
they do not appear useful for distinguishing failing and healthy firms over the
short run.
For net operating asset turnover the coefficients are positive, however, and
highly significant in the fourth and fifth years prior to failure. The strength of
significance in the earlier years is consistent with the expectation that the effi-
ciency variables would be more important over the long run. The positive sign
on the coefficient indicates that the failing firms have higher net operating asset
turnover than healthy firms. It is not likely that this results from higher sales,
but these firms may be neglecting investment in capital assets, making them no
longer able to compete with firms more vigorous in investing, and so more likely
to fail. It may be that net book value of assets based on historical cost might be
noisy because of lack of homogeneity of depreciation methods across one com-
pany’s fixed asset structure and also a lack of homogeneity of methods across
companies included in the sample.
Both receivables turnover and inventory turnover have predictable positive
FAILURE OF US 1NL)USlRIAL FIRMS
ASSESSING VULNERABILI1‘Y 1‘0 41
signs for their coefficients. Inventory turnover provides an especially strong
indicator in the long run. Not so receivables turnover, however; it becomes
significant at a level greater than 90 per cent in years one, two and four.
The acid test ratio is particularly important over the short run, as an inade-
quate reserve of quick assets could precipitate bankruptcy. This is shown by the
sign of its coefficients on this ratio as well as its high significance (at the 99 per
cent confidence level) for years one to three.
Current assets to total assets is significant in the first year prior to failure
while the sign of its coefficient is not. A very high significance (97 per cent) and
a positive sign calculated for this variable in the fifth year indicates that a
relatively high investment in fixed capital distinguished doomed firms from
those with bright prospects.
Return on invested capital appears significant only for year four in the logit
estimation, indicating that accounting measured profits do not distinguish fail-
ing from healthy firms. This unexpected result could be due to several factors:
(1) the profitability of failing and healthy firms really does not differ; (2) profit
on accounting reports is subject to alternative accounting treatments (these
might result in measurement error which would bias the coefficients and T-
statistics downward (Bohrnstedt and Carter, 1971, p.131); or (3) the profit
measures are ‘managed’ figures subject to the choice of accounting principles.
Clever accounting can make the profits of failing firms appear to be as high as
those of healthy firms. Indeed, Ohlson found that many of the failing firms
which his model missed showed profits at least as high as those of healthy firms;
many even showed a profit during their year of failure (Ohlson, p. 130).

Evaluation of the Models for Classificalion Accuracy on In-Sample Data and Predicliue
AccuraGy on Out-of-sample Data
Applying the estimated models to the data for each firm yields a probability of
failure conditional on that firm’s financial attribute vector. A firm from the
initial sample will be classified as ‘failed’ or ‘healthy’ according to whether its
predicted probability falls above or below a critical probability. The selection of
the relevant probability determines to a great extent the classification results.
This selection determines the optimal tradeoff between either the probability of
Type 1 error, (failing to identify a failing firm), and the probability of Type 2
error, (failing to identify a healthy firm). If the costs of these error types are the
same, then the cutoff probability should balance these error rates. In this case,
a cutoff probability which minimizes total errors would be appropriate. Ohlson
(p.124) and others imply as much, especially when they choose not to break
error rates into these categories or minimize total errors.
For optimal classification or prediction, a cut-off probability must reflect the
differential in relative costs of Type 1 and Type 2 errors for each decision set-
ting. A commercial bank loan decision, for example, involves calculating the
potential cost of a Type 1 error in extending a loan to a failing client. The bank
might lose interest as well as principal, whereas in making a Type 2 error the
42 ZAVCREN

bank would lose a n opportunity to earn money on the loan. Diamond (1976,
pp. 127 - 32) estimated that the ratio ofType 1 cost to Type 2 cost was in a range
of 20/1 to 38/1. Altman, el al. used survey data from small Southeast regional
banks to estimate a Type 1 to Type 2 error ratio of 35/1 (Altman, Haldman and
Narayanan, pp.44-6). These proportions counsel caution, but a bank still has
a great responsibility to be accurate. T h e cost of a Type 2 error would be higher
to the firm’s management than to a commercial bank. T h e firm could unfairly
lose reputation and credit as the error in labelling became a self-fulfilling pro-
phesy.
Since both error types are important, and little is known about the true costs
of Type 1 and Type 2 errors, a cut-off probability which minimizes the total
error rate was used. ’I’hisyields an estimated classification error rate computed
on the same basis as those reported in other studies and hence they are directly
comparable. the minimum total classification error rates for years one to five
prior to bankruptcy were 18 per cent, 17 per cent, 28 per cent, 27 per cent and
20 per cent, respectively. T h e error rate for one year prior to bankruptcy was
similar to Ohlson’s (the only year he reported). T h e error rates for the earlier
years are similar to (or slightly lower than) the average error rates reported in
other studies, and significantly lower than that reported by Altman.
For the purpose of assessing how well the results aid generalization, it is
necessary to evaluate the models on some sample other than that for which they
were originally estimated. Due to the limited availability of data for bankrupt
firms, this is seldom done on a truly distinct sample. Instead substitute pro-
cedures are frequently used, such as an (n - 1) holdout estimation evaluation
on the same sample only. All such procedures bias error rate estimation
significantly downward as is discussed in Eisenbeis (1977). T o avoid this bias,
the predictive accuracy of the models in this study was evaluated using a sample
of failed firms from years after the original test period.
Data was available on 16 New York Stock Exchange firms which failed in
1979 and 1980. These firms were matched with healthy firms by the same pro-
cedure as used in the original estimation. T h e coefficients for each of the five
years’ models were applied to data for each firm for the appropriate year.
Predictive accuracy was assessed using the same total error rate criterion as was
used above. The resulting error rates were 31 per cent for years one through
five, respectively.
There is little basis for comparison with other studies, as only Beaver and
Blum use a true holdout sample to evaluate the predictive success of their
models. Their holdout samples draw from the same time period as their estima-
tion period; this would bias error rates downward. It would be better to assess
the ability to generalize the results over a different time period. Thus, both
studies report lower error rates than they would have if inter-temporal
generalization were sought, as is analyzed in Zavgren (1983).
The question of generalizability is a n important issue, since due to data
limitations none of these studies can be based on random sampling techniques.
ASSESSING VULNEKABILI'I'Y '1'0 FAILURE OF U S INDUSTRIAL FIRMS 43
The results presented in this study cover the population (albeit a small one) of
firms failing from 1972 to 1978 for which there is adequate data. Therefore, the
ability to proceed from sample to a generalization about population for this
time period is not an issue. The only question about generalization is whether
the model results can be validly extended to a later time period. Thus, the more
stringent validation test was selected for this study.

SUMMARY A N D CONCLUSIONS

This study has developed logit and probit models of bankruptcy to generate a
probability of failure as a financial risk measure, and to test the pattern of
significance of the financial attributes in the models over a five year period
prior to failure. Models which generated a probability of failure as a cardinal
measure of risk proved to be more useful than the dichotomous classification
usually obtained from discriminant analysis models. The latter turns out to be
too stringent a partition of the outcome space for most decision settings. T h e
models estimated here were found to be highly significant (at greater than the 99
per cent confidence level) in distinguishing between failing and healthy firms
over the five-year period.
The information content of the models was evaluated using information-
theoretic measures. The models for the earliest years prior to failure contained
an amount of information which compares favorably with Theil's results. l'he
amount of information over the subsequent five-year period increases by an
average of 18 per cent for the failed firms and 16 per cent for the non-failed
firms.
Classification and prediction error rates were also evaluated for the models.
They were found to compare favorably with other models, especially for
prediction ability when the stringency of the inter-temporal generalizability
test is considered.
The significance of the coefficients for each of the variables in the models
were traced for each of the five years. The pattern of significance was found to
be highly congruent with a prion'expectations. The efficiency ratios were found
to have the most significance over the long run, which indicated that efficiency
in the utilization of assets is difficult to modify over the short run. Profitability
was not found to be a significant distinguishing characteristic. The negative
coefficient and high significance of the acid test ratio in later years would
indicate that ability to meet current obligations is a very important factor in
avoiding bankruptcy. The coefficients of the liquidity measure in earlier years
and its negative sign indicate that the failing firms were more interested in
liquidity than productive opportunities. Debt proved to be a significant
characteristic and was consistently higher for ailing than for healthy firms.
Financial ratios can provide highly significant measures for evaluating
bankruptcy risk. In addition, the pattern of significance of the coefficients in
44 ZAVGREN

these models indicates that these variables would be important for helping a
manager or analyst to assess risk.

REFERENCES

Altman, E.I. (1968), ‘Financial Ratios, Discriminant Analysis and the Prediction of Corporate
Bankruptcy,’ TheJournal ofFinance (September 1968), pp. 589-609.
-, R.G. Haldman and P. Narayan (1977), ‘Zeta Analysis,’Journal ofBanking and Finance,
(June 1977), pp. 29-54.
Beaver, W.H. (1965), ‘Financial Ratios as Predictors of Failure’ (Unpublished Ph.D.
dissertation, Graduate School of Business, University of Chicago, 1965).
-(1966), ‘Financial Ratios as Predictors of Failure,’ Empirical Research in Accounfing:
SelecfedSfdies, supplmunf to Journal ofAccounling Research (1966), pp. 71 - 11 1 .
Blum, M. (1974). ‘Failing Company Discriminant Analysis,’ Journal of Accounfing Research
(Spring 1974), pp. 1-25.
Bohrnstedt, G.W. and T . Carter (1971), ‘Robustnes in Regression Analysis,’ Sociological
Mefhodoloo ed. H.L. Costner, (Jossey-Bass, San Francisco 1971), pp. 118- 145.
Chen. K.H. and T.A. Shimerda (1980), ‘An Empirical Analysis of useful Financial Ratios,’
FinancialManagmf (Spring 1981), pp. 51 - 60.
Chesser, D. L. (1974), ‘Predicting Loan Noncompliance,’ The Journal of Commercial Eank Lending
(August 1974), pp. 28-38.
Deakin, E.B. (1972), ‘A Discriminant Analysis of Predictors of Business Failure,’ Journal of
Accounting Research (Spring 1972), pp. 167- 179.
Diamond, H.S., Jr. (1976), ‘Pattern Recognition and the Detection of Corporate Failure’
(Unpublished Ph.D dissertation, New York University, 1976).
Edmister, R . O . (1972), ‘An Empirical Test of Financial Ratio Analysis for Small Business
Failure Prediction,’ Journal of Financial and Quanfifatiuc Analysis (March 1972), pp.
1477- 1493.
Eisenbeis, R.A. (1977), ‘Pitfalls in the Application of Discriminant Analysis in Business and
Economics,’ Thc,Journal ofFinance (June 1977), pp. 875-900.
F and S Index of Corporafe Changes (1972 - 78), (Predicasts, Inc., Cleveland 1972 - 78).
Finney, D.J., Probit Analysis, second edition, (Cambridge University Press, New York, 1952).
Korobow, L. and D.P. Stuhr, (1975), ‘Toward Early Warning of Changes in Banks’ Financial
Condition: A Progress Report,’ Fcdnal Resnue of New York Monfhly Rcuiew (1975), pp.
157- 165.
Lev, B. and S. Sunder (1979), ‘Methodological issues in the Use of Financial Ratios,’Journal of
Accounting and Economics (Winter, 1979), pp. 187-210.
McFadden, D. (1973), ‘Conditional Logit Analysis of Qualitative Choice Behavior,’ Fronfiers in
Economefrics, ed. P. Zarembka (Academic Press, 1973), pp. 105 - 142.
Maddala, G.S. (1977), Economefrics(McGraw-Hill, 1977).
Martin, D. (1977), ‘Early Warnings of Bank Failure,’Journal of Banking and Finance, pp. 249-67.
Meyer, P.A. and H.W. Pifer (1970). ‘Prediction of Bank Failures,’Journal ofFinancc (September,
1970). pp. 853-868.
Modigliani, F. and M.H. Miller, “l’he Cost oICapital, Corporation Finance, and the ’l‘heory of
Investment,’ The American Economic Reuiew, (June 1958), pp. 453-487.
Morrison, G.D. (1972), ‘Upper Bounds for Correlation Between Binary Outcomes and
Probabilistic Predictions,’ Journal of fhc American SfafisficalAssociation (March 1972). pp.
68-70.
Ohlson, J.A. (1980). ‘Financial Ratios and the Probabilistic Prediction of Bankruptcy,’.Journal of
Accounfinf Research (Spring 1980), pp. 109- 131.
Pinches, G.E., K.A. Mingo and J.K. Caruthers (1973), ‘The Stability of Financial Patterns in
Industrial Organizations,’Journal ofFinance (May 1973), pp. 389-396.
-_ _ ___ and A.A. Eubank (1975), ‘The Hierarchical Classification of Financial
Ratios,’Journal ofBusiness Research (October 1975), pp. 295-310.
Pindyck, R.S. and D.L. Rubinfeld (1976), EconomefricModrlsand EconomicForecasfs(McGraw-Hill,
1976).
ASSESSING VULNERABILITY TO FAILURE O F US INOUSTRIAL FIRMS 45
Santomero, A. and J.D. Vinso (1977), ‘Estimating the Probability of Failure for Commercial
Banks and the Banking System,’Journal of Banking and Finance (September, 1977), pp.
185-205.
Sinkey, J.F., Jr. (1975), ‘A Multivariate Statistical Analysis of the Characteristics of Problem
Banks,’ The Journal of Finance (March, 1975), pp. 21 -36.
Smith, V. (1972). ‘Default Risk, Scale, and the Home-Made Leverage Theorem,’ American
Economic Review, (February, 1972), pp. 66-76.
Stiglitz, J.E. (1972), ‘Some Aspects of the Pure Theory of Corporate Finance: Bankruptcies and
Take-Overs,’ The Bell Journal of Economics and Managcment Science (Autumn, 1972), pp.
458 - 482.
Theil, H . (1971), Principles oJEconometrics (John Wiley and Sons, 1971).
Tinsley, P.A. (1970), ‘Capital Structure, Precautionary Balances, and Valuation of the Firm: the
Problem of Financial Risk,’Journal of Financial and Quantitative Anahsis (March, 1970). pp.
33-62.
White, K.J. (1979), S H A Z A M : ‘An Econometrics Computer Program’, Version 3.0.I , Rice University
Social Science Computer Lab, Houston (November, 1979).
White, R.W. and M . Turnbull (1975). ‘The Probability of Bankruptcy: American Railroads’
(Working paper, Institute of Finance and Accounting, London University Graduate
School of Business, July 1975).
Wilcox, J.W. (1971), ‘A Simple Theory of Financial Ratios as Predictors of Failure,’Journaf of
Accounting Research (Autumn 1971). pp. 389-395.
-(1973), ‘A Prediction of Business Failure Using Accounting Data,’ Empirical Research in
Accounting, Selected Studies, Supplement to Vol. 11, Journal ofAccounting Research (1973). pp.
163-79.
White, R.W. and M . Turnbull (1975), ‘The Probability of Bankruptcy for American Industrial
Firms’ (Working Paper, Institute of Finance and Accounting, London University
Graduate School of Business, July 1975).
Zavgren, C.V. (1983). ‘The Prediction of Corporate Failure: The State of the Art’ Journal of
Accounting Litcrafure (Spring, 1983), pp. 1-38.
Zmijewski, M.E. (1983). ‘Predicting Corporate Bankruptcy: An Empirical Comparison of the
Extant Financial Distress Models’ (Working Paper, State University of New York at
Buffalo, 1983).

You might also like