Anda di halaman 1dari 12

Journal of Contemporary Accounting & Economics 6 (2010) 34–45

Contents lists available at ScienceDirect

Journal of Contemporary
Accounting & Economics
journal homepage: www.elsevier.com/locate/jcae

Research Note

A comparison of alternative bankruptcy prediction models


Y. Wu *, C. Gaunt, S. Gray
Clive Gaunt, Stephen Gray, UQ Business School, University of Queensland, Australia

a r t i c l e i n f o a b s t r a c t

Article history: Early models of bankruptcy prediction employed financial ratios drawn from pre-bank-
Received 16 April 2009 ruptcy financial statements and performed well both in-sample and out-of-sample. Since
Revised 16 November 2009 then there has been an ongoing effort in the literature to develop models with even greater
Accepted 16 November 2009
predictive performance. A significant innovation in the literature was the introduction into
Available online 3 April 2010
bankruptcy prediction models of capital market data such as excess stock returns and stock
return volatility, along with the application of the Black–Scholes–Merton option-pricing
Keyword:
model. In this note, we test five key bankruptcy models from the literature using an up-
Bankruptcy prediction models
to-date data set and find that they each contain unique information regarding the proba-
bility of bankruptcy but that their performance varies over time. We build a new model
comprising key variables from each of the five models and add a new variable that proxies
for the degree of diversification within the firm. The degree of diversification is shown to be
negatively associated with the risk of bankruptcy. This more general model outperforms
the existing models in a variety of in-sample and out-of-sample tests.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction

Since Beaver (1966), a large literature on bankruptcy prediction has emerged, and its influence has spilled into the com-
mercial world, where it has been used in the development of several commercially employed bankruptcy prediction models.
Out of this literature have come a number of competing empirical models with alternative explanatory variables and alter-
native statistical methodologies for model estimation.
The dependent variable in these models is commonly a dichotomous variable where ‘‘firm filed for bankruptcy” is set to 1
and ‘firm remains solvent’ is set to 0. The independent variables are often accounting ratios drawn from financial statements
and include measures of profitability, liquidity, and leverage. Some studies also include market-driven variables such as the
volatility of stock returns and past excess returns. The accounting-based models developed by Altman (1968) and Ohlson
(1980) have emerged as the most popular bankruptcy prediction models and are often used by empirical accounting
researchers as indicators of financial distress.1
Altman (1968) employs multivariate discriminate analysis (MDA) on a list of financial ratios to identify those ratios that
are statistically associated with future bankruptcy. Ohlson (1980) uses a logit model, which uses less restrictive assumptions
than those taken by the MDA approach. Zmijewski (1984) adopts a probit approach that is also based on accounting data but
uses a different set of independent variables. All of these approaches predict future bankruptcy based on accounting ratios
drawn from publicly available financial statements.
More recently, Shumway (2001) has proposed a discrete-time hazard model to predict a firm’s bankruptcy using both
accounting and market variables. The main difference between this model and the static logit model is that the hazard model

* Address: UQ Business School, University of Queensland, Brisbane 4072, Australia. Tel.: +61 (0) 733 469 318.
E-mail address: y.wu@business.uq.edu.au (Y. Wu).
1
See, for example, Dichev (1998), Stone (1991), Francis (1990) and Burgstahler et al. (1989).

1815-5669/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jcae.2010.04.002
Y. Wu et al. / Journal of Contemporary Accounting & Economics 6 (2010) 34–45 35

can be estimated within the logit framework while using the entire life span of information (all firm-years) for each firm. By
contrast, the static logit model can only incorporate one firm-year for each observation (i.e., each observation consists of a
single set of variables observed at a single point in time).
Another stream of the bankruptcy prediction literature focuses on market-based information. Among others, Hillegeist
et al. (2004) have developed a BSM-Prob bankruptcy prediction model that is based on the Black–Scholes–Merton option-
pricing model. Their results indicate that the BSM-Prob model outperforms the models of Altman (1968) and Ohlson
(1980) in a series of tests.
There are also a number of papers that propose various firm-characteristics that may be useful additional predictors of
future bankruptcy. For example, Rose (1992) proposes a model of firm diversification in which managers use diversification
to reduce the risk of bankruptcy, particularly where the ratio of the manager’s firm-specific human capital to his non-firm-
specific human capital is high. Denis et al. (1997) measure corporate diversification by the number of business segments.
Beaver et al. (2005) propose that, other things equal, large firms have a smaller probability of bankruptcy and that a part
of this explanation is related to corporate diversification. That is, corporate diversification and firm-size are two firm-char-
acteristics that may help to predict future bankruptcy.
Hillegeist et al. (2004) compare the performance of their BSM-Prob model against the Altman and Ohlson models in a
series of in-sample and out-of-sample tests, concluding that the BSM model outperforms the accounting-based models. Sim-
ilarly, Chava and Jarrow (2004) examine the relative performance of Shumway’s hazard model against the Altman and Zmi-
jewski models, concluding that the hazard model outperforms static logit models.
In this note, we make three contributions to the bankruptcy prediction literature. First, we note that the existing liter-
ature has compared the relative performance of various subsets of models. We perform a more comprehensive examination
of the performance of models that use different data (accounting, market, and firm-characteristic data) and different econo-
metric approaches (MDA, logit, hazard, and BSM models). Second, we develop a number of additional performance metrics
in various tables and figures, which provide some new perspectives on the relative performance of the various approaches.
Third, we examine an integrated model that encompasses accounting variables, market information (including the BSM
probabilities from the model proposed by Hillegeist et al.) and firm-characteristics such as size and corporate
diversification.
The extant literature reports that the predictive ability of each model varies over time, and our empirical analysis con-
firms this result.2 For example, the Ohlson model performs relatively well in the mid- to late-1980s, whereas the Shumway
model performs better in the 1990s. This time-series variation in the performance of different models that use different data
and employ different econometric techniques suggests that each model might be capturing slightly different aspects of corpo-
rate financial health. These points towards an integrated model that includes accounting data, market data, and firm-character-
istics such as size and corporate diversification. We develop such a combined model and show that it outperforms the various
existing models in a range of in-sample and out-of-sample forecasting tests. We conclude that each model does indeed capture
slightly different aspects of corporate financial health and that a combined model is able to provide more consistent and more
reliable indications of corporate bankruptcies over a broader range of market conditions.

2. Models and data

2.1. Models

There are a number of key models that have been developed by various authors and presented in the bankruptcy predic-
tion literature over the last three decades. These are:

(i) Altman (1968) – MDA model based on accounting variables.


(ii) Ohlson (1980) – logit model with accounting ratios.
(iii) Zmijewski (1984) – probit model using accounting data.
(iv) Shumway (2001) – hazard model with both accounting and market variables; and
(v) Hillegeist et al. (2004) – BSM-Prob model based on both accounting and market variables.

Using an up-to-date data set, we compare the performance of these models and propose and test a new, more general,
model that includes accounting information and various firm-characteristic variables including size and corporate diversi-
fication. Table 1 provides a brief summary of each model.

2.2. Data

Data on firm bankruptcies is obtained from New Generation Research (www.bankruptcydata.com), Compustat and CRSP.
The sample contains NYSE- and AMEX-listed Compustat firms and covers the period from 1980 to 2006. Consistent with the
previous literature, a firm is defined as bankrupt if the firm makes a chapter 11 filing within 1 year. The accounting data (i.e.,

2
A more detailed analysis of model performance in various sub-periods, not reported in the main tables, is available upon request.
36 Y. Wu et al. / Journal of Contemporary Accounting & Economics 6 (2010) 34–45

Table 1
Summary of empirical models and variables employed. This table sets out the various bankruptcy prediction models that we examine. The first column lists the
models that are examined. The second column summarizes the model specification. The final columns document the explanatory variables that are used in the
models.

Model Formula Variable Description


0
Altman (1968) Z¼bX X1 = Net working capital/total assets
Multiple-discriminant where Z is the MDA score and X represents the X2 = Retained earnings/total assets
analysis variables listed
Cutoff point: X3 = Earnings before interest and taxes/total assets.
Z P 2.675, classified as non-bankrupt X4 = Market value of equity/book value of total
liabilities.
Z < 2.675, classified as bankrupt X5 Sales/total assets.
Ohlson (1980) 0
P ¼ ð1 þ expfb XgÞ 1 Ohlsonsize = Log(total assets/GNP price-level index). The
Logit model where P is the probability of bankruptcy and X index assumes a base value of 100 for 1968.
represents the variables listed. The logit function TLTA = Total liabilities divided by total assets.
maps the value of b0 X to a probability bounded WCTA = Working capital divided by total assets.
between 0 and 1. CLCA = Current liabilities divided by current assets.
OENEG = 1 If total liabilities exceed total assets, 0
otherwise.
NITA = Net income divided by total assets.
FUTL = Funds provided by operations (income from
operation after depreciation) divided by total
liabilities.
INTWO = 1 If net income was negative for the last 2 years,
0 otherwise.
CHIN = (NIt  NIt1)/(|NIt| + |NIt1|), where NIt is net
income for the most recent period. The
denominator acts as a level indicator. The variable
is thus intended to measure the relative change in
net income.
Zmijewski (1984) P ¼ Uðb0 XÞ NITL = Net income divided by total liabilities.
Probit model where P is the probability of bankruptcy and X TLTA = Total liabilities divided by total assets.
represents the variables listed, and UðÞ CACL = Current assets divided by current liabilities.
represents the cumulative normal distribution
function. The probit function maps the value of
b0 X to a probability bounded between 0 and 1.
Shumway (2001) P i;t ¼ ð1 þ expfyi;t gÞ1 NITL = Net income divided by total liabilities.
Hazard model 2 3 TLTA = Total liabilities divided by total assets.
X 1;t1 :::X 1;tj
yi;t ¼ a þ b0 X i;t1 ¼ b0 4 ::::: 5 Relative Size = Log(the number of outstanding shares
X n;t1 :::X n;tj multiplied by year-end share price then divided
where P is the probability of bankruptcy and X by total market value).
represents the variables listed. This is a logit LagExReturn = Cumulative annual return in year t  1 minus
model, but instead of treating each firm-year as the value-weighted CRSP NYSE/AMEX index
an independent observation, all prior values of the return in year t  1.
independent variables for a particular firm are LagSIGMA = Standard deviation of the residual derived from
included in the information set. n Represents the regressing monthly stock return on market return
number of independent variables, and j represents in year t  1.
the number of time periods prior to time t for
which data are available.
 
Hillegeist et al. (2004) P¼N 
lnðV A =XÞþðld0:5rA2 ÞT
pffiffi VE = Current market value of equity.
Black–Scholes option- rA T VA = Current market value of assets.
pricing model where P is the probability of bankruptcy and the X = Face value of debt maturing at time T.
variables are as defined. d = Continuous dividend rate expressed in terms of
VA.
rA ¼ rE V E =ðV E þ XÞ:
l = Continuously compounded expected return on
assets.
T = Debt maturity, set as 1 year.
Multi-period logit model P i;t ¼ ð1 þ expfyi;t gÞ1 EBITTA = Earnings before interest and taxes/total assets.
with expanded set of 2 3 WCTA = Working capital divided by total assets.
X 1;t1 :::X 1;tj
variables yi;t ¼ a þ b0 X i;t1 ¼ b0 4 ::::: 5 TLMTA = Total liabilities divided by market value of total
Hazard model with X n;t1 :::X n;tj assets.
additional explanatory where P is the probability of bankruptcy and X CHIN = (NIt  NIt1)/(|NIt| + |NIt1|), where NIt is net
variables represents the variables listed. n Represents the income for the most recent period. The
number of independent variables and j represents denominator acts as a level indicator. The variable
the number of time periods prior to time t for is thus intended to measure the relative change in
which data are available. net income.
Excess = Cumulative annual return in year t  1 minus
Return the value-weighted CRSP NYSE/AMEX index
return in year t  1.
Y. Wu et al. / Journal of Contemporary Accounting & Economics 6 (2010) 34–45 37

Table 1 (continued)

Model Formula Variable Description


Sigma = Standard deviation of the residual derived from
regressing monthly stock return on market return
in year t  1.
Price = Log(closing price of previous fiscal year).
Segment = Diversification factor, which is proxied by the
number of business segments.

financial ratios and diversification data) are obtained from Compustat, and the daily and monthly stock returns are collected
from CRSP.
We assume that annual financial statements are available by the end of the fourth month after the firm’s fiscal year-end.
For each bankrupt firm, the most recent financial statements must be available at least 4 months prior to bankruptcy. If the
bankruptcy is within 4 months, we treat the previous year’s financial statements as the last available observation. The infor-
mation will be less informative if the lag between recent available annual financial statements and the bankruptcy filing date
is too long, so we exclude bankruptcies with more than 18 months between the most recent financial statements and the
bankruptcy filing date. To be consistent with previous studies (see, for example, Begley et al., 1996, and Shumway, 2001),
we only include firms with a Standard Industrial Classification (SIC) code that is less than 4000 or between 5000 and
5999 (effectively, we exclude financial firms).
The number of bankrupt companies with complete accounting data available is 887. Among them, 650 firms have a bank-
ruptcy filing date available in the New Generation Research database. For those bankrupt companies without an exact filing
date, we treat the delisting date from CRSP or their deletion date from Compustat, whichever is earlier, as the bankruptcy
date. All estimations are on a yearly basis. We exclude firms that are delisted from CRSP or deleted from Compustat for rea-
sons other than bankruptcy or liquidation. A total of 737 firm-years are excluded due to missing values. The final sample for
estimation consists of 50,611 firm-years. This data set contains 887 bankruptcies and 49,724 non-bankrupt firm-years.
We employ both matched-pair and random-sampling methods to ensure comparability with the estimation procedures
of the models under investigation. Similar to Altman’s (1968) procedure, we construct a matched-pair based on firm-size and
the two-digit SIC industry classification code on a yearly basis. We measure firm-size by total book assets. A match is based
on minimizing the absolute value of the ratio of the difference between the size of the bankrupt firm and that of the healthy
firm to the size of the bankrupt firm.
We also construct a new, more general, model based on a set of selected variables, which includes the sorts of profitabil-
ity, liquidity, and leverage variables that have appeared in the previous literature. Earnings before interest and tax to total
assets (EBITTA) and change in income (CHIN) are two powerful measures that capture the static and dynamic condition of a
firm’s profitability condition. Working capital to total assets (WCTA) is the most widely used measure of liquidity, and its
explanatory power is evident in the existing bankruptcy literature. We also use a market-based leverage measure, total lia-
bilities to the market value of total assets (TLMTA), which is proposed by Campbell et al. (2008). They find that a market-
based measure of leverage performs better that the traditional book value to leverage ratio. They suggest that market prices
may more rapidly incorporate new information about the firm’s prospects. We also include the lagged stock return (lagEx-
Return) and volatility variables (lagSIGMA) used in the Shumway model. To avoid multicollinearity problems, we only in-
clude one firm-size variable, the log price per share (PRICE) in our model. In addition, we include a variable to measure
the extent to which the firm’s operations are diversified. Our Segment variable measures the number of separate business
segments in each of our sample firms.
Table 2 presents summary statistics for each variable used in this study over the period 1980–2006. Our data set contains
887 bankruptcies and 49,724 non-bankrupt firm-years, which are defined as ‘‘healthy” in the table. For each variable, we
display the means in bold if they are significantly different at the 5% level between the two sub-samples. To remove the im-
pact of extreme values and outliers, all variables have been winsorized at the 1st and 99th percentiles.
Table 2 groups the variables into a number of categories. The profitability variables measure the ability of the firm to gen-
erate sufficient profits or returns to remain a viable going concern. EBIT to total assets (EBITTA) and net income to total assets
(NITA) both measure the profitability of the firm in relation to its asset base. For these measures, our sample of firms that
became bankrupt in the following year score substantially less than the remainder of our sample.
We follow Ohlson (1980) in measuring change in net income as CHIN = (NIt  NIt1)/(|NIt| + |NIt1|) where NIt represents
net income in year t. Where net income was negative in the previous year but positive in the current year, this variable will
take the value 1. Where the opposite occurs, the variables will the value 1. These events occur for a number of observations
in our sample, so we focus more on the median than on the mean for this variable. In this regard, we note that the median
bankrupt firm sees its net income deteriorate by 36% in the year prior to bankruptcy, whereas net income grows by 6% for the
other firms in our sample.
38 Y. Wu et al. / Journal of Contemporary Accounting & Economics 6 (2010) 34–45

The sales to total assets variable (SALES) was originally used by Altman (1968). Other things equal, it would be expected
that firms with a higher value of this variable would be less likely to experience bankruptcy. However, Table 2 indicates that
the reverse is true. Further analysis reveals that the book value of total assets for the mean bankrupt firm decreased by 9.8%
over the year prior to bankruptcy.3 Consequently, the ratio of sales over the course of the year to total assets at the end of the
year can be misleading. We retain this variable in our analysis of the Altman model, but we do not use it in our proposed com-
prehensive model.
The liquidity variables measure the ability of the firm to meet its short-term obligations. On average, our sample of bank-
rupt firms has lower working capital to total assets (WCTA). The bankrupt firms also have higher current liabilities relative to
current assets (CLCA)4 and are more likely to have negative net income over the prior 2 years (INTWO). Bankrupt firms are also
likely to have a lower ratio of funds from operations to total liabilities (FUTL).
The leverage variables measure the relative amount of debt and other obligations of the firm. On average, the bankrupt
firms have lower retained earnings relative to total assets (RETA), a lower market value of equity to total liabilities (METL),
and higher total liabilities relative to total assets (TLTA). The bankrupt firms are also more likely to report total liabilities in
excess of the book value of total assets (OENEG) and to have higher liabilities relative to the market value of total assets
(TLMTA).
The firm-size variables are drawn from the literature where, other things equal, larger firms are thought to be more able
to trade through difficult times and are less likely to become bankrupt. Consistent with the findings in the existing literature,
the firms that become bankrupt in the following year are, on average, smaller. We include the log of the firm’s share price
(PRICE) as a firm-size variable, as share prices tend to be related to firm-size with larger firms, on average, commanding a
higher price per share.5 We also measure size as total assets relative to a GNP price index (Ohlsonsize) for the implementation
of the Ohlson model and as the firm’s market capitalization relative to the entire market (Relative Size) as this is required for the
Shumway model. The share price may also reflect liquidity, with the literature documenting that very low-priced stocks exhibit
lower liquidity on average. Table 2 indicates that our sample of bankrupt firms, on average, has lower stock prices than the
remainder of our sample.
The last panel of Table 2 contains a number of other firm-characteristics. On average, our sample of bankrupt firms tends
to have lower and more volatile stock returns (lagSIGMA) than the other firms in our sample. Our measure of excess returns
(lagExReturn) is the CRSP total return minus the CRSP value-weighted-index total return. The mean and median bankrupt
firm has exhibited substantial underperformance in the year prior to bankruptcy. The bankrupt firms also tend to be younger
(logage) and have fewer business segments (Segment), indicating a lower degree of corporate diversification.

3. Methodology and results

3.1. Explanatory power of alternative models in a nested logit framework

The various bankruptcy prediction models that have been proposed differ in two ways – they use different explanatory
variables and different econometric techniques (MDA, probit, and logit). We begin by examining the role of the explanatory
variables; while holding the econometric technique constant, we use the logit framework outlined in Shumway (2001). In
the subsequent sub-section we also vary econometric techniques.
Specifically, we begin by testing the explanatory power of each model with a multi-period logit approach. Initially, all
tests are conducted in-sample. This procedure generates a series of nested models, which allows for model comparison
and significance tests among different sets of variables. The multi-period logit model is equivalent to the hazard model in
Shumway (2001), except that we only include one firm-year observation for each bankrupt firm but all firm-year observa-
tions for the surviving firms. This refinement significantly improves model performance. As there are serially correlated
observations included in the model (because the data set contains more than 1 year of data for some firms), the standard
error will be biased upwards. Consequently, standard errors are adjusted according to Shumway (2001), where the Wald
Chi-Square generated by the logit program is divided by the average number of firm-years.
Table 3 reports the parameter estimates from each model. Here we only test the variables used in four accounting-based
models. The BSM-Prob model is not based on accounting ratios and therefore does not fit within this framework. In all cases,
we use the same set of variables as the original authors, but we employ the logit framework of Shumway (2001). The esti-
mated coefficients are in bold type when they are statistically significant at the 5% level. The adjusted Wald chi-squared sta-
tistics are computed as in Shumway (2001). The dependent variable is set to 1 for bankrupt firms and 0 otherwise, so a
positive coefficient indicates that a higher value of the relevant variable increases the likelihood of bankruptcy.
All of the models that are examined use a range of profitability, liquidity, and leverage variables. In general, and as ex-
pected, a firm is more likely to become bankrupt if it has lower profitability, lower liquidity and higher leverage. The majority
of the variables in these categories are statistically significant. Some models also include one of the firm-size variables. Ta-
ble 3 indicates that, other things equal, smaller firms are more likely to experience bankruptcy.

3
A table of summary statistics for all accounting information is available upon request.
4
Even though the sample has been winsorized, there remain some firms with very low current assets, resulting in high values of the ratio of current
liabilities to current assets, and this affects the mean estimate. We note that the median value of this variable is also much weaker for the bankrupt firms.
5
A detailed correlation matrix is available upon request.
Table 2
Summary statistics for explanatory variables. This table reports summary statistics for all of the variables used in this study over the period 1980–2006. Our final data set contains 887 bankruptcies and 49,724 non-
bankrupt firm-years (defined as ‘‘healthy” in the table). Mean values are in bold if the difference between the bankrupt and healthy samples is significant at the 5% level. EBITTA = earnings before interest and taxes to
total assets; SALES = sales to total assets; NITA = net income divided by total assets; CHIN = (NIt  NIt1)/(|NIt| + |NIt1|), where NIt is net income for the most recent period. WCTA = working capital to total assets;
CLCA = current liabilities to current assets; FUTL = income from operations after depreciation divided by total liabilities; INTWO = 1 if net income was negative for the previous 2 years, 0 otherwise. RETA = retained
earnings to total assets; METL = market equity to total liabilities; TLTA = total liabilities to total assets; OENEG = 1 if total liabilities exceeds total assets, 0 otherwise; TLMTA = total liabilities to market value of total
assets. Ohlsonsize = log(total assets/GNP price-level index), the index assumes a base value of 100 for 1968; Relative Size = log(the number of outstanding shares multiplied by year-end share price then divided by

Y. Wu et al. / Journal of Contemporary Accounting & Economics 6 (2010) 34–45


total market value); price = log of closing price at end of previous fiscal year. LagSIGMA = historical idiosyncratic risk; LagExReturn = lagged excess return; logage = log(years for which firm has traded); Segment = the
number of business segments in the firm.

Profitability variables Liquidity variables


EBITTA SALES NITA CHIN WCTA CLCA FUTL INTWO
Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy
Mean 0.18 0.03 1.35 1.22 0.31 0.02 0.33 0.02 0.03 0.3 2.81 0.68 0.79 0.01 0.59 0.19
Standard deviation 0.32 0.21 0.96 0.8 0.41 0.23 0.6 0.55 0.28 0.23 33.2 6.12 12.95 1.38 0.49 0.39
Maximum 0.25 0.32 4.48 4.48 0.24 0.24 1 1 0.81 0.85 979 925 2.81 15.4 1 1
Median 0.06 0.08 1.19 1.09 0.15 0.04 0.36 0.06 0.05 0.28 0.88 0.47 0.07 0.14 1 0
Minimum 1.15 1.15 0 0 1.46 1.46 1 1 0.43 0.43 0.01 0 385 74.88 0 0
Leverage variables Firm-size variables
RETA METL TLTA OENEG TLMTA Ohlsonsize Relative Size Price
Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy
Mean 1.04 0.18 1.43 6.33 0.79 0.47 0.23 0.02 0.69 0.34 4.54 3.64 12.59 10.34 0.49 2.37
Standard deviation 2.05 1.3 4.81 11.99 0.28 0.23 0.42 0.12 0.25 0.23 1.75 2.05 1.64 2.12 1.31 1.19
Maximum 0.83 0.83 67.02 79.33 1.23 1.23 1 1 0.93 0.93 2.06 4.16 5.33 5.33 4.42 4.42
Median 0.24 0.17 0.28 2.25 0.8 0.47 0 0 0.78 0.31 4.58 3.79 12.67 10.48 0.63 2.55
Minimum 8.31 8.31 0.08 0.08 0.05 0.05 0 0 0.01 0.01 9.65 10.76 15.28 15.28 2.08 2.08

Other firm-characteristics
LagSIGMA LagExReturn Logage Segment
Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy Bankrupt Healthy
Mean 0.21 0.14 0.44 0.09 2.2 2.41 2.14 3.37
Standard deviation 0.11 0.09 0.5 0.69 0.77 0.82 2.42 3.49
Maximum 0.56 0.56 3.29 3.29 4.23 4.23 20.00 42.00
Median 0.19 0.12 0.57 0.04 2.2 2.4 1.00 2.00
Minimum 0.03 0.03 0.95 0.95 1.1 1.1 1.00 1.00

39
40 Y. Wu et al. / Journal of Contemporary Accounting & Economics 6 (2010) 34–45

The Shumway model and our proposed comprehensive model also include a range of other firm-characteristics. Table 3
indicates that, other things equal, a firm is more likely to become bankrupt if it has recently experienced negative excess
returns (i.e., the stock price has fallen after adjusting for the market) (lagExReturn) and if stock price returns are more vol-
atile (lagSIGMA). These results are consistent with Shumway (2001). Table 3 also indicates that the age of the firm (logage)
has no significant explanatory power.
Our proposed comprehensive model includes the sorts of profitability, liquidity, and leverage variables that have ap-
peared in the previous literature. It also includes the lagged return and volatility variables used in the Shumway model.
To avoid multicollinearity problems, we only include one firm-size variable (PRICE) in our model. In addition, we include
a variable to measure the extent to which the firm’s operations are diversified (Segment). Table 3 indicates that more diver-
sified firms (i.e., those with a greater number of business segments) are less likely to become bankrupt, other things being
equal.
We also compare each of the models based on model-fit criteria such as the Receiver Operating Characteristics area (ROC
area) and Pseudo R2. The ROC curve is a widely used technique for assessing various rating methods in the bankruptcy pre-
diction literature (see, for example, Chava and Jarrow, 2004; Vassalou and Xing, 2004; Agarwal and Taffler, 2008).
The ROC score and Pseudo R2 statistics are higher for the proposed comprehensive model than for any others that are
examined. The ROC score is discussed in Agarwal and Taffler (2008) and Lyandres and Zhdanov (2007). A scaled version
of this score, known as the accuracy ratio, is used by Moody’s and is discussed at some length in Vassalou and Xing
(2004). A perfect model has an ROC score of 1, and a model that has no discriminatory power whatsoever has an ROC score
of 0.5. Thus, a higher ROC score indicates that the model is better able to discriminate between bankrupt and healthy firms.
The comprehensive model has an RPC score of 0.93 and a Pseudo R2 of 36%, which is superior to all other models that are
examined.
The in-sample logit analysis in Table 3 indicates that the profitability, liquidity, and leverage variables in our proposed
comprehensive model are all statistically significant. Other things equal, a firm is more likely to become bankrupt if it
has lower profitability, lower liquidity and higher leverage. The additional explanatory variables improve the performance
of the model and are jointly significant (a likelihood ratio test of the joint significance of the other variables is significant at
the 1% level). In particular, the lagged excess stock return, the volatility of stock returns, the share price (as a proxy for firm-
size, and perhaps liquidity) and firm diversification are individually and jointly significant.
In the remainder of this section, we compare this proposed new model against other models from the literature in a num-
ber of more formal specification and performance tests. That is, having identified the set of accounting, market, and firm-
characteristic variables, we next turn to variations in the econometric specification of the model.

3.2. Original econometric specifications

In this section, we re-estimate each of the models using the econometric specification used by the original authors. We
thus use multiple-discriminant analysis (MDA) for the Altman model, standard logit for the Ohlson model and probit for the
Zmijewski model. The Shumway and comprehensive models are estimated in a logit-hazard setting, and the BSM model is
estimated as described in Hillegeist et al. (2004). For all models, we use all available observations. For example, MDA is used
to estimate the Altman model, but it is based on all observations in our sample rather than on the matched-pairs for the
bankrupt firms as in the original paper.
The result of this procedure is an estimated bankruptcy probability from each model for each observation in our sample.
We rank these model bankruptcy probabilities from lowest to highest. For a given percentile, we then report the frequency of
Type I (classifying a bankrupt firm as healthy) and Type II (classifying a healthy firm as bankrupt) errors. For example, at the
50th percentile we consider the incidence of classification errors if all observations with a model bankruptcy probability
above the 50th percentile are classified as bankrupt and all others are classified as healthy.
In Table 4, we report the rate of Types I and II errors. For each model, we use bold font to report the point at which the
sum of Types I and II classification errors is minimized. This procedure is a modified version of that used by Begley et al.
(1996).6
Table 4 shows that for all models the sum of the classification errors is minimized with a cutoff point between the 75th
and 90th percentiles. For our proposed comprehensive model, for example, total classification error rates are minimized by
classifying as bankrupt only those firms with a bankruptcy probability above the 87th percentile. We note that the classifi-
cation error rate for the comprehensive model is lower than that of all other models.

6
Begley et al. (1996) report results based on the bankruptcy probability from the Ohlson model rather than based on percentiles, as we have done. In their
sample, the sum of classification errors is minimized using an estimated bankruptcy probability of 3.8%. That is, any firm with an estimated probability of
bankruptcy more than 3.8% should be classified as bankrupt. Note that this cannot be directly compared with our percentile ranking, which is not in terms of
raw probabilities but rather in terms of the ranking of this probability relative to the rest of the sample. Of course, minimizing the sum of Type I and Type II
errors is not necessarily optimal. In some circumstances, one type of error might be more costly than another. For example, a corporate regulator seeking
advance warning of impending corporate collapses would be more concerned with Type I errors (bankrupt firms being classified as healthy). This might lead the
regulator to set a lower cutoff value to reduce the incidence of Type I errors at the expense of more Type II errors (having to conduct more detailed evaluations
of firms that turn out to be healthy).
Y. Wu et al. / Journal of Contemporary Accounting & Economics 6 (2010) 34–45 41

Table 3
Estimation results. This table presents the parameter estimates from various bankruptcy prediction models. Our sample is from 1980 to 2006 and contains 887
bankruptcies and 49,724 non-bankrupt firm-year observations. Bold font signifies an estimate that is statistically significant at the 5% level. The adjusted Wald
Chi-Square statistics from the logistic regression are presented adjacent to each parameter estimate. The ROC statistic measures the ability of a model to
discriminate between bankrupt and healthy firms, with a higher score indicating a better ability. ***,**, and * indicate statistical significance at the 1%, 5%, and
10% levels, respectively.

Altman (1968) Ohlson (1980) Zmijewski (1984) Shumway (2001) Comprehensive model
Estimate Wald Chi- Estimate Wald Chi- Estimate Wald Chi- Estimate Wald Chi- Estimate Wald Chi-
Square Square Square Square Square
Intercept 3.2*** 216.15 7.2*** 162.25 5.81*** 155.69 11.71*** 149.23 5.79*** 104.05
Profitability variables
EBITTA 3.27 34.93 1.75*** 13.52
SALES 0.18*** 7.27
NITA 0.03 0.01 1.47*** 34.93 0.33 1.17
CHIN 1.02*** 24.16 0.44** 5.31
Liquidity variables
WCTA 3.38*** 47.19 1.87*** 13.85 1.69*** 12.69
CLCA* 0 0 0.28** 5.38
FUTL 0.06 1.33
INTWO 1.16*** 19.98
Leverage variables
RETA 0.1 1.04
METL 0.24*** 13.98
TLTA 3.69*** 32.35 3.54*** 56.5 3.71*** 80.23
OENEG 0.54 1.63
TLMTA 4.44*** 52.97
Size Variables
Ohlsonsize 0.17*** 7.63
Relative Size 0.4*** 31.89
PRICE 0.34*** 9.57
Other firm-characteristics
LagExReturn 1.33*** 21.59 0.54* 3.77
LagSIGMA 2.52** 5.26 2.67** 5.2
Logage 0.04 0.08
Segment 0.18*** 10.34
2 Log- 7125 6623 7188 6284 5705
likelihood
ROC score 0.861 0.887 0.852 0.906 0.929
Pseudo R2 0.202 0.259 0.195 0.297 0.361
*
Zmijewski (1984) uses current assets to current liabilities, whereas Ohlson (1980) uses the reciprocal.

3.3. Out-of-sample performance of alternative models with rolling windows

Our next set of tests examines the out-of-sample performance of the various models. We conduct a series of rolling out-
of-sample estimations similar to that of Hillegeist et al.(2004). For example, the first estimation is based on firm-year obser-
vations from 1980 and bankruptcies in 1981. The estimated coefficients are then used to predict bankruptcies in 1982 with
data up to 1981. The second estimation is expanded to use firm-year observations from 1980 to 1981 and bankruptcies in
1982. The second set of estimated coefficients is then used to predict bankruptcies in 1983 with data up to 1982. The win-
dows continue expanding; the estimated coefficients used to predict bankruptcies in 2006 are based on firm-year observa-
tions from 1980 to 2004 and bankruptcies from 1981 to 2005.
For the BSM model, all observations are out-of-sample in the sense that no coefficients need to be estimated. Rather,
parameter estimates based on observable data are used for all observations. Although we could therefore consider the entire
sample to be an out-of-sample test of the BSM model, we restrict our analysis to the same out-of-sample period that we use
for the other models. We do this to guard against the possibility that it may have been easier or more difficult to distinguish
between bankrupt and healthy firms in the estimation period relative to the out-of-sample period. This would give the BSM
model an unfair advantage or disadvantage, as the case may be.
The out-of-sample performance of each model is summarized in Table 5. For each model, we begin by estimating the
bankruptcy probability for each observation in the out-of-sample period. This is based on coefficients estimated from the
earlier data period. We then rank all observations in the out-of-sample period by the estimated bankruptcy probability,
and we group observations into deciles on this basis. Decile 1 represents those firms with the highest estimated probability
of bankruptcy.
We then report the actual number of bankruptcies from among those firms classified in each decile. There were 874 ac-
tual bankruptcies in our out-of-sample period. A perfect model would have all of these ranked in the top decile. Thus, a better
42 Y. Wu et al. / Journal of Contemporary Accounting & Economics 6 (2010) 34–45

Table 4
Bankruptcy classification rates. This table shows the incidence of Type I errors (classifying a bankrupt firm as healthy) and Type II errors (classifying a healthy
firm as bankrupt) according to model scores. For each model, the cutoff that minimizes the sum of Types I and II errors is highlighted in bold. For example,
under the Altman model if companies with model scores above the 85th percentile are classified as bankrupt, and those with model scores below the 85th
percentile are classified as healthy, Types I and II error rates are 64.8% and 15.6%, respectively. *For comparison, we use the negative Altman’s Z score here.

Model score Altman Ohlson Zmijewski Shumway BSM Comprehensive


(percentile) Type I Type II Type I Type II Type I Type II Type I Type II Type I Type II Type I Type II
50 0.348 0.507 0.041 0.502 0.062 0.502 0.564 0.501 0.513 0.502 0.535 0.501
70 0.492 0.306 0.103 0.299 0.122 0.299 0.072 0.298 0.136 0.300 0.049 0.298
71 0.497 0.296 0.106 0.289 0.126 0.289 0.077 0.288 0.145 0.290 0.053 0.288
72 0.504 0.286 0.108 0.279 0.131 0.279 0.080 0.278 0.153 0.280 0.057 0.278
73 0.516 0.276 0.115 0.269 0.142 0.269 0.084 0.268 0.163 0.270 0.058 0.268
74 0.522 0.266 0.118 0.258 0.146 0.259 0.085 0.258 0.172 0.260 0.062 0.257
75 0.527 0.256 0.125 0.248 0.149 0.249 0.091 0.248 0.180 0.249 0.070 0.247
76 0.538 0.246 0.127 0.238 0.157 0.239 0.092 0.238 0.190 0.239 0.075 0.237
77 0.546 0.236 0.133 0.228 0.161 0.229 0.096 0.228 0.199 0.229 0.079 0.227
78 0.565 0.226 0.136 0.218 0.165 0.219 0.102 0.217 0.205 0.219 0.080 0.217
79 0.584 0.216 0.146 0.208 0.172 0.209 0.107 0.207 0.213 0.209 0.083 0.207
80 0.595 0.206 0.153 0.198 0.180 0.199 0.111 0.197 0.222 0.199 0.085 0.197
81 0.603 0.196 0.160 0.188 0.191 0.189 0.117 0.187 0.230 0.189 0.089 0.187
82 0.615 0.186 0.165 0.178 0.201 0.179 0.125 0.177 0.241 0.179 0.092 0.177
83 0.626 0.176 0.173 0.168 0.220 0.169 0.133 0.167 0.252 0.169 0.096 0.166
84 0.637 0.166 0.179 0.158 0.233 0.159 0.144 0.157 0.260 0.159 0.103 0.156
85 0.648 0.156 0.187 0.148 0.247 0.149 0.154 0.147 0.276 0.149 0.114 0.146
86 0.661 0.146 0.198 0.138 0.253 0.139 0.168 0.137 0.293 0.140 0.121 0.136
87 0.673 0.136 0.209 0.128 0.264 0.129 0.183 0.127 0.298 0.129 0.130 0.126
88 0.688 0.127 0.220 0.118 0.274 0.119 0.202 0.117 0.322 0.120 0.142 0.116
89 0.699 0.117 0.238 0.108 0.289 0.109 0.218 0.108 0.350 0.110 0.156 0.106
90 0.707 0.107 0.256 0.098 0.312 0.099 0.232 0.098 0.367 0.100 0.163 0.096
99 0.870 0.018 0.687 0.014 0.760 0.016 0.641 0.014 0.763 0.016 0.547 0.012

model has more of the actual bankruptcies classified in the top decile(s) and fewer classified in the bottom decile(s). Table 5
reports the number and proportion of the actual bankruptcies that were classified into each of the deciles by the estimated
bankruptcy probability. It also reports the proportion of firms in each decile that did experience bankruptcy.
For our proposed comprehensive model, for example, 683 of the 874 out-of-sample bankruptcies (78%) are classified into
Decile 1. The top two deciles contain 89% of the out-of-sample bankruptcies. This performance is substantially better than all
of the other models.7
Table 5 also reports a model performance rating for each of the models. This score is based on a weighted-average of the
proportion of bankruptcies in each decile. A model that classifies all of the actual out-of-sample bankruptcies into Decile 1
performs perfectly and receives a rating of 100%. By contrast, a model that classifies all of the actual out-of-sample bankrupt-
cies into Decile 10 is actively incorrect and receives a rating of 100%. A model that allocates the actual bankruptcies evenly
across all 10 deciles is equivalent to a random assignment and receives a rating of 0. On this dimension, the comprehensive
model also performs better than all of the other models that are examined.

3.4. Information content tests

We perform two types of information content tests similar to those performed by Hillegeist et al. (2004). First, we esti-
mate a series of discrete hazard models – one for each of our bankruptcy prediction models. In each case, we include the
lagged rate of bankruptcies (the proportion of total sample firms that experienced bankruptcy during the previous year)
as a proxy for the baseline hazard rate. We then include the score (fitted prediction value) from one of the bankruptcy pre-
diction models and examine the relative explanatory power of the baseline rate and the prediction model.
The results are reported in Panel A of Table 6. In all cases, a positive coefficient indicates that the relevant variable is asso-
ciated with an increase in the likelihood of bankruptcy. For all models, the coefficient on the model score is positive and sig-
nificant at the 5% level. This indicates that the model has explanatory power beyond the baseline lagged bankruptcy rate.
That is, the variables used in the model and the econometric method for combining them into a single bankruptcy score
are useful in predicting bankruptcies – beyond knowing the bankruptcy rate from the previous year. This result is consistent
with Hillegeist et al. (2004).
The results in Table 6 also indicate that the lagged bankruptcy rate is generally not significant in combination with the
bankruptcy scores from the various models. This is in contrast with the results of Hillegeist et al., who find that the coeffi-

7
For the BSM Model, a small number of observations are lost due to the non-convergence of the estimation algorithm or missing data. In an unreported table,
we generate results from a sample that is restricted to those observations for which the BSM model can be estimated, and the overall ranking is unchanged.
Y. Wu et al. / Journal of Contemporary Accounting & Economics 6 (2010) 34–45 43

Table 5
Out-of-sample forecast accuracy with rolling windows. This table reports statistics relating to out-of-sample performance based on the rolling windows
method. The details of the rolling windows method are in the text. We classify observations into deciles based on these bankruptcy probabilities (Decile 1 being
the highest). Except for the BSM model, each decile contains 4953 observations, 874 of which are bankruptcies (for the BSM model, 727 of which are
bankruptcies). For each model, we report the actual number of bankruptcies in each decile, the proportion of the 874 actual bankruptcies that occur in each
decile, and the proportion of the total firms in each decile that experienced bankruptcy. We also report a ‘‘model performance rating”: where all bankruptcies
occur in Decile 1, the model performs perfectly and the score is 100%; where all bankruptcies occur in Decile 10, the model is actively incorrect and the score is
100%; where bankruptcies are evenly spread across deciles, the score is 0%, and so on.

Decile 1 2 3 4 5 6 7 8 9 10
Altman model
Number of actual bankruptcies in decile 247 100 94 69 59 74 73 51 56 51
Proportion of bankruptcies in each decile 0.283 0.114 0.108 0.079 0.068 0.085 0.084 0.058 0.064 0.058
Proportion of firms that experience bankruptcy 0.050 0.020 0.019 0.014 0.012 0.015 0.015 0.010 0.011 0.010
Model performance rating 28.73%
Ohlson model
Number of actual bankruptcies in decile 518 180 72 37 21 22 12 5 2 5
Proportion of bankruptcies in each decile 0.593 0.206 0.082 0.042 0.024 0.025 0.014 0.006 0.002 0.006
Proportion of firms that experience bankruptcy 0.105 0.036 0.015 0.007 0.004 0.004 0.002 0.001 0.000 0.001
Model performance rating 79.74%
Zmijewski model
Number of actual bankruptcies in decile 511 193 58 35 26 23 8 5 7 8
Proportion of bankruptcies in each decile 0.585 0.221 0.066 0.040 0.030 0.026 0.009 0.006 0.008 0.009
Proportion of firms that experience bankruptcy 0.103 0.039 0.012 0.007 0.005 0.005 0.002 0.001 0.001 0.002
Model performance rating 78.54%
Shumway model
Number of actual bankruptcies in decile 528 127 76 40 27 15 23 11 16 11
Proportion of bankruptcies in each decile 0.604 0.145 0.087 0.046 0.031 0.017 0.026 0.013 0.018 0.013
Proportion of firms that experience bankruptcy 0.107 0.026 0.015 0.008 0.005 0.003 0.005 0.002 0.003 0.002
Model performance rating 73.96%
BSM model
Number of actual bankruptcies in decile 412 135 74 30 25 20 8 6 6 11
Proportion of bankruptcies in each decile 0.567 0.186 0.102 0.041 0.034 0.028 0.011 0.008 0.008 0.015
Proportion of firms that experience bankruptcy 0.095 0.031 0.017 0.007 0.006 0.005 0.002 0.001 0.001 0.003
Model performance rating 75.24%
Comprehensive model
Number of actual bankruptcies in decile 683 94 39 23 13 15 0 2 3 2
Proportion of bankruptcies in each decile 0.781 0.108 0.045 0.026 0.015 0.017 0.000 0.002 0.003 0.002
Proportion of firms that experience bankruptcy 0.138 0.019 0.008 0.005 0.003 0.003 0.000 0.000 0.001 0.000
Model performance rating 89.22%

cient on the lagged bankruptcy rate is positive and statistically significant for the models and sample that they examine.
They conclude that a higher lagged bankruptcy rate is indicative of a generally higher probability of bankruptcy in the cur-
rent year, one that is over and above the bankruptcy probability from the model being examined. We find no strong relation-
ship between current and lagged bankruptcy probabilities in our sample. We further examine this in a model that includes
only the lagged bankruptcy probability and no other explanatory variables. In this case, the coefficient is again positive, but it
fails to reach statistical significance and has a Pseudo R2 statistic less than 1%. The likely reason for this is the fact that our
sample contains all observations for which data are available. This is in contrast with previous papers that include a rela-
tively small sample of non-bankrupt firms. The fact that our sample contains many more non-bankrupt firms means that
the bankruptcy rates in our sample are very small for all years. Although there are more bankruptcies during recessions
and market downturns, the variation in the bankruptcy rate over time is small, making it harder for this variable to reach
statistical significance. The information content test shows that the bankruptcy score generated from our comprehensive
model has the highest explanatory power as indicated by the log-likelihood ratio and Pseudo R2 statistics. The significance
persists even when we combine all bankruptcy scores in one regression.
Panel B contains a set of Vuong (1989) tests comparing the relative performance of pairs of models. Hillegeist et al. (2004)
argue that prediction-oriented tests cannot determine whether the model performance differences are statistically signifi-
cant, while relative information content tests can. The aim of the relative information content tests is to examine whether
one variable (or a set of variables) contains more information than another variable (or set of variables). The test is designed
to compare the amount of bankruptcy-related information in each model. Following Hillegeist et al. (2004), a discrete hazard
model is used to test for relative information content and the log-likelihood statistics for each case are compared using the
Vuong (1989) logit-based test.
In each case, a positive test statistic indicates that the ‘‘row” model outperforms the ‘‘column” model and a negative test
statistic indicates that the reverse is true. All of the statistics in the table are significant at the 5% level. The negative figures in
the right-hand column indicate that our proposed comprehensive model significantly outperforms all other candidate
models.
44
Table 6
Information content tests. Panel A reports relative information content tests as described in Hillegeist et al. (2004). In each case, we estimate a logit regression with an intercept, the lagged bankruptcy rate (the

Y. Wu et al. / Journal of Contemporary Accounting & Economics 6 (2010) 34–45


proportion of total firms that experienced bankruptcy in the previous year – to proxy for economy-wide factors), and the score from one of the bankruptcy prediction models. The models are estimated on the full
sample from 1980 to 2006. Wald chi-squared statistics are adjusted according to Shumway (2001), and coefficients are reported in bold if they are significant at the 5% level. The final columns report a combined model
that contains the scores from all individual models. Panel B reports Vuong test statistics. Negative values indicate that the ‘‘column model” beats the ‘‘row model.” For example, the Ohlson model beats the Altman
model. All of the Vuong test statistics are significant at the 1% level, where the critical value is 2.33. ***,**, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.

Altman Ohlson Zmijewski Shumway BSM Comprehensive Combined


Parameter Estimate Wald {2 Estimate Wald {2 Estimate Wald {2 Estimate Wald {2 Estimate Wald {2 Estimate Wald {2 Estimate Wald {2
Panel A: information content tests
Intercept 3.93*** 184.16 0.07 0.04 0.00 0.00 0.08 0.05 4.92*** 239.73 0.89*** 6.89 0.63 1.56
Lagged bankruptcy rate 24.70 1.09 8.19 0.11 3.12 0.02 5.14 0.04 38.63 2.31 28.95 1.23 23.31 0.73
Z score (Altman) 0.24*** 16.91 0.04 0.60
O score (Ohlson) 1.02*** 197.13 0.22 1.97
ZM score (Zmijewski) 1.02*** 140.13 -0.02 0.03
S score (Shumway) 1.00*** 202.6 0.06 0.13
BSM score (Hillegeist et al.) 1.21*** 123.07 0.07 0.28
New score 0.91*** 205.03 0.74*** 36.46
2 Log-likelihood 7337.98 5526.85 5921.62 5272.11 6110.67 4781.67 4741.38
Pseudo R2 0.02 0.26 0.21 0.30 0.19 0.36 0.37

Ohlson Zmijewski Shumway BSM Comprehensive


Panel B: Vuong test statistics
Altman 166.10*** 138.59*** 164.26*** 110.45*** 198.09***
Ohlson 33.60*** 43.01*** 13.80*** 111.62***
Zmijewski 7.80*** 4.19*** 63.92***
Shumway 12.14*** 81.10***
BSM 59.59***
Y. Wu et al. / Journal of Contemporary Accounting & Economics 6 (2010) 34–45 45

4. Summary and conclusion

In this paper, we examine the empirical performance of a number of bankruptcy prediction models. The models use a
range of different independent variables (accounting information, market and firm-characteristic data, and implied proba-
bilities from option-pricing models) and a range of different econometric specifications (multiple-discriminant analysis, lo-
git, probit, and hazard models).
We find that the key accounting information relates to profitability, liquidity, and leverage. Specifically, firms are more
likely to experience bankruptcy if they have relatively lower earnings before interest and tax to total assets, a larger decline
in net income, relatively low working capital to total assets, or high market-based leverage – total liabilities to the market
value of total assets.
We also find that a number of firm-characteristic variables assist in forecasting bankruptcies. Firms are more likely to
experience bankruptcy if the lagged stock returns are large and negative or the lagged volatility is relatively high. Smaller
firms and firms with fewer business segments are also more likely to experience bankruptcy, other things equal. Firms with
a higher implied probability of bankruptcy (estimated in relation to an option-pricing model) are also more likely to expe-
rience bankruptcy, other things equal.
In this paper, we examine the empirical performance of a range of bankruptcy prediction models using a series of in-sam-
ple and out-of-sample performance metrics. We examine a number of metrics used in the extant literature and develop some
new and modified performance metrics. Across all of these tests, both in-sample and out-of-sample, we find that a compre-
hensive model that includes key accounting information, market data, and firm-characteristics significantly outperforms
models from the extant literature.
Specifically, the MDA model of Altman (1968) performs poorly relative to other models in the literature. The accounting-
based models of Ohlson (1980) and Zmijewski (1984) perform adequately during the 1970 but their performance has dete-
riorated over more recent periods. The hazard model of Shumway (2001), which includes market data and firm-character-
istics, generally outperforms models that are based on accounting information only.8 The use of option-implied probability
proposed by Hillegeist et al. (2004) performs adequately but is generally inferior to the Shumway model. We conclude that
a more comprehensive model that draws inferences from key accounting information, market data, and firm-characteristics
provides the most reliable forecasts of future bankruptcy. This is consistent with the different types of data capturing different
aspects of corporate financial distress.

Acknowledgements

The Australian Research Council provided funds for this project under Grant DP0663048. We thank Stephen Hillegeist and
Elizabeth Keating for sharing programs to estimate Black–Scholes probabilities and Vuong test statistics.

References

Agarwal, Vineet, Richard Taffler, 2008. Comparing the performance of market-based and accounting-based bankrutpcy prediction models. Journal of
Banking & Finance 32, 1541–1551.
Altman, Edward.I., 1968. Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. Journal of Finance 23, 589–609.
Beaver, William H., 1966. Financial ratios as predictors of failure. Journal of Accounting Research, 71–111.
Beaver, William H., McNichols, Maureen F., Rhie, Jung-Wu, 2005. Have financial statements become less informative? Evidence from the ability of financial
ratios to predict bankruptcy. Review of Accounting Studies 10, 93–122.
Begley, Joy, Ming, Jin, Watts, Susan, 1996. Bankruptcy classification errors in the 1980s: an empirical analysis of Altman’s and Ohlson’s models. Review of
Accounting Studies 1, 267–284.
Burgstahler, David, Jiambalvo, James, Noreen, Eric, 1989. Changes in the probability of bankruptcy and equity value. Journal of Accounting and Economics
11, 207–224.
Campbell, John Y., Hilscher, Jens, Szilagyi, Jan, 2008. In search of distress risk. Journal of Finance 63, 2899–2939.
Chava, Sudheer, Jarrow, Robert A., 2004. Bankruptcy prediction with industry effects. Review of Finance 8, 537–569.
Denis, David J., Diane K., Denis, Atulya Sarin, 1997. Agency problems, equity ownership, and corporate diversification. Journal of Finance 52, 135–160.
Dichev, Ilia D., 1998. Is the risk of bankruptcy a systematic risk? Journal of Finance 53, 1131–1147.
Francis, Jennifer, 1990. Corporate compliance with debt covenants. Journal of Accounting and Research 28, 326–347.
Hillegeist, Stephen A., Elizabeth Keating, Donald P., Cram, Jyle G., Lunstedt, 2004. Assessing the probability of bankruptcy. Review of Accounting Studies 9,
5–34.
Lyandres, Evgeny, Alexei Zhdanov, 2007. Investment opportunities and bankruptcies prediction. Working Paper, Rice University, Houston.
Ohlson, James, 1980. Financial ratios and the probabilistic prediction of bankruptcy. Journal of Accounting Research 18, 109–131.
Rose, David, 1992. Bankruptcy risk, firm-specific managerial human capital, and diversification. Review of Industrial Organisation 7, 65–73.
Shumway, Tyler, 2001. Forecasting bankruptcy more accurately: a simple hazard model. Journal of Business 74, 101–124.
Stone, Mary, 1991. Firm financial stress and pension plan continuation/replacement decisions. Journal of Accounting and Public Policy 3, 175–206.
Vassalou, Maria, Yuhang Xing, 2004. Default risk in equity returns. Journal of Finance 59, 831–868.
Vuong, Quang H., 1989. Likelihood ratio tests for model selection and non-nested hypotheses. Econometrica 57, 307–333.
Zmijewski, Mark E., 1984. Methodological issues related to the estimation of financial distress prediction models. Journal of Accounting Research 22
(Suppl.), 59–82.

8
An exception are tests of out-of-sample forecast accuracy using the rolling windows method.