Journal of Empirical Finance 17 (2010) 818833

Contents lists available at ScienceDirect

Journal of Empirical Finance

j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / j e m p f i n

A hybrid bankruptcy prediction model with dynamic loadings on

accounting-ratio-based and market-based information: A binary quantile
regression approach
Ming-Yuan Leon Li a,1, Peter Miu b,
a
Department of Accountancy and Graduate Institute of Finance and Banking, College of Management, National Cheng Kung University,
No. 1, University Rd., Tainan, Taiwan
b
DeGroote School of Business, McMaster University, 1280 Main Street West, Hamilton, ON, Canada L8S 4M4

a r t i c l e

i n f o

Article history:
Received 23 April 2009
Received in revised form 24 March 2010
Accepted 2 April 2010
Available online 18 April 2010
JEL classication:
G33
C51

Keywords:
Binary quantile regression
z-score
Distance-to-default
Bankruptcy

a b s t r a c t
While using the binary quantile regression (BQR) model, we establish a hybrid bankruptcy
prediction model with dynamic loadings for both the accounting-ratio-based and marketbased information. Using the proposed model, we conduct an empirical study on a dataset
comprising of default events during the period from 1996 to 2006. In this study, those rms
experienced bankruptcy/liquidation events as dened by the Compustat database are classied
as defaulted rms, whereas all other rms listed in the Fortune 500 with over a B-rating
during the same time period are identied as survived rms. The empirical ndings of this
study are consistent with the following notions. The distance-to-default (DD) variable derived
from the market-based model is statistically signicant in explaining the observed default
events, particularly of those rms with relatively poor credit quality (i.e., high credit risk).
Conversely, the z-score obtained with the accounting-ratio-based approach is statistically
signicant in predicting bankruptcies of rms of relatively good credit quality (i.e., low credit
risk). In-sample and out-of-sample bankruptcy prediction tests demonstrated the superior
performance of utilizing dynamic loadings rather than constant loadings derived by the
conventional logit model.
2010 Elsevier B.V. All rights reserved.

1. Introduction
The difculty in predicting corporate failure has posed a long standing problem in credit risk research. The most famous credit risk
model is Altman's (1968) z-score model, which employs accounting-ratio-based information in determining and quantifying how the
probability of default and a set of nancial ratios are related. The Altman's z-score model relies mostly on information obtained from
companies' nancial statements; whereas a different approach is based on the work by Merton (1974) which suggests that one can use
a company's debt ratio together with its asset value volatility to predict its default probability.2 This market-based approach serves as
the building block of a number of credit risk models commonly used in practice (e.g. that of Moody's KMV).
This study differs from previous related works by offering a new perspective on the link between Altman's and Merton's
models. In particular, this study explores the effectiveness of a hybrid model, in which information from both accounting-ratio-

Corresponding author. Tel.: +1 905 525 9140x23981.

E-mail addresses: [email protected] (M.-Y.L. Li), [email protected] (P. Miu).
1
Tel.: +886 6 2757575x53421; fax: +886 6 2744104.
2
In particular, Merton (1974) considers that holding the debt of a risky company is equivalent to holding the debt of a risk-free company plus being short a put
option on the assets of the company. The put option arises because if the value of the assets falls below the value of the debt, the shareholders can put the assets to
debt holders, and in return, receive the right not to repay the full amount of the debt. In this analogy, the underlying for the put option is the company assets, and
the strike price is the amount of debt.
0927-5398/$ see front matter 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.jempn.2010.04.004

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

819

based and market-based approaches is incorporated. This approach merits detailed study. In combining the two sets of
information in a hybrid model, we may enhance the predictive power of a company's default event if both the accounting-ratio
variables involved in Altman's model and the equity-based information obtained with Merton's model convey company-specic
credit risk information that is not subsumed by each other.3
In establishing a hybrid model, one key issue is the determination of the optimal loading for each of the two types of
information when they are incorporated in the model. In this study, we examine a bankruptcy prediction model utilizing dynamic
loadings established by the binary quantile regression (BQR). It can be considered as a generalized version of those hybrid models
with static loadings, which have been examined in previous studies.
Our ideas are presented below. Typically, in developing a credit risk model, nancial institutions conduct statistical analyses
with the historical data of various characteristics of both the defaulted and survived borrowers and their loans. The loading for
each borrower/loan characteristic is estimated by ensuring the observed default events could be best explained. Using the vector
of calibrated loadings, nancial institutions can therefore decide on whether to approve or reject a loan application by knowing the
borrower's characteristics which can in turn be used in assessing her credit quality and default probability. A key question
addressed in this study is whether the optimal loading of each borrower/loan characteristics is different between good and poor
credit quality companies. If they are different, should banks attach more weight to accounting-ratio-based content or marketbased information when poor credit quality companies are concerned? Such dynamic assignment of loadings may enhance the
overall predictive power of the credit risk model over loan portfolios of different credit qualities.
A number of studies have pointed out that investors and nancial institutions rarely opt for only one approach, but rather
combine different sources of information in arriving at their own credit risk assessments. In particular, most closely related is the
research by Miller (1998), Kealhofer and Kurbat (2001), Kealhofer (2003), Lfer (2007) and Mitchell and Roy (2008). Except for
the ndings of Kealhofer and Kurbat (2001), these studies conclude that combining various failure prediction models improves the
prediction of default over the use of a single measure.4 However, in determining the weights to be assigned to the various default
prediction techniques, these studies either employ the straightforward logit regression or use some subjective combination rules.5
The predictive power of these hybrid models is potentially limited owing to the use of a constant or an exogenously imposed
loading for each credit score or specic ranges of each credit score.
In practice, nancial institutions recognize the benets of adopting a risk rating system which allows for different weighting
schemes for different credit portfolios. For example, given that market-based information is more readily available and thus more
reliable for large corporate borrowers than mid-market ones, banks tend to attach a heavier weight to market-based information
when they assess the credit risks of the former than the latter. However, any such segmentations are likely to be exogenously
imposed (e.g. by size, geography, etc.) rather than driven by statistical analyses with the objective of ensuring minimal default
prediction errors.
Unlike previous studies, we are one of the rst to employ dynamic loadings in a bankruptcy prediction model of which not only
the magnitudes of the loadings but also their applicability over the population of interest are endogenously determined. By
utilizing dynamic loadings established by BQR, we allow the data to reveal to us the implicit segmentations that can ensure the
least prediction errors over the whole dataset. In this setup, the segmentations and the optimal loadings on the credit-related
explanatory variables are jointly determined. Specically, we consider a BQR model in which these loadings vary with the quantile
levels of credit risks of the borrowers.6 We also conduct an empirical study to demonstrate the benet of our proposed dynamic
model in explaining the observed default events in comparison with an alternative hybrid model utilizing constant loadings.7
The rest of this paper is organized as follows. Section 2 presents the literature review. In Section 3, we describe the hybrid credit
risk models used in this study, including (1) the setting with constant loadings generated by the conventional logit regression, and
(2) the setting with dynamic loadings under the BQR model. The model evaluation methods are dened in Section 4. Section 5
presents the data sources and empirical results. Finally, a few concluding remarks are made in Section 6.

2. Literature review
Under the regulatory capital requirements of Basel II, banks are allowed to develop their own credit risk models in adopting the
internal ratings-based approach in determining the capital charges with respect to the credit risks of their portfolios. Research in
this area has generated considerable interest. In this study, we examine how accounting-ratio-based and market-based
information may be used in a joint fashion in implementing such models.

In practice, nancial institutions typically use both types of information in assessing the credit quality of their corporate clients in their risk raking systems.
Kealhofer and Kurbat (2001) examine the default prediction power of Merton's approach relative to expert debt ratings and accounting variables and show
that agency credit ratings have no incremental value for default prediction.
5
For example, a subjective combination rule is specied under the standardized approach of Basel II. It species that banks working with two credit ratings
must assign the higher (i.e., maximum) risk assessments to its borrowers; whereas banks working with three credit ratings or more must use the highest of the
two lowest risk assessments (see Basel Committee, 2006).
6
In this study, the credit risk quantile of a borrower is dened as the percentage of borrowers which are expected to have a lower credit risk (based on the
credit model) than the borrower being considered.
7
Unlike Lfer (2007) in which separate loadings are estimated for companies with low vs. high credit score being subjectively dened based on, for example,
their Moody's ratings, we let the data reveal the optimal segmentation based on the interaction of all the explanatory variables and the observed default events
utilizing BQR. We believe our approach is more exible and thus can ensure our conclusions to be more robust to any potential misclassication errors.
4

820

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

Accounting-ratio-based models are typically built by using a set of accounting-ratio variables. The most famous of such models
is Altman's z-score. Following Altman's (1968) study, Mensah (1984) indicates that the past performance involved in a rm's
accounting statements may not be informative in predicting the future, and thus suggests that it is necessary to regenerate the
accounting-ratio-based models periodically. Hillegeist et al. (2004) argue that the ability of accounting information in predicting
bankruptcy is likely to be limited given the fact that they are formulated to describe the nancial condition of the company under
the going-concern principle (i.e., assuming it will not go bankrupt).
Based on the criticisms of accounting-ratio-based models, market-based models are proposed by Black and Scholes (1973) and
Merton (1974). It is claimed that market prices reect future expected cash ows, and thus should be more useful in predicting
bankruptcy. Market-based models are further examined by a number of studies, including Hillegeist et al. (2004), Reisz and Perlich
(2004), Vassalou and Xing (2004) and Campbell et al. (2006), in assessing default probability.
Research comparing market-based and accounting-ratio-based bankruptcy prediction models has also been conducted.
Although Altman's (1968) z-score model may suffer from a lack of theoretical underpinning, the validity of Merton's model is also
limited by a number of stringent assumptions (see for example Saunders and Allen, 2002). It is therefore not surprising that the
empirical evidence on the relative performance of market-based against accounting-ratio-based models is mixed (see Kealhofer,
2003; Oderda et al., 2003; Hillegeist et al., 2004; Reisz and Perlich, 2004; Stein, 2005; Campbell et al., 2006; Blochlinger and
Leippold, 2006 and Agarwal and Tafer, 2008).
Even if one model is superior to the other, it does not imply that the inferior model should be neglected altogether. It may be
possible to combine the two models to form an even better one (see Miller, 1998; Kealhofer and Kurbat, 2001; Kealhofer, 2003;
Lfer, 2007; Mitchell and Roy, 2008). This study claims that both accounting-ratio-based and market-based information should
be valuable for bankruptcy prediction. Both types of information are considered as credit risk indicators and are utilized
simultaneously in explaining the probability of default. Notably, the weights assigned to the two types of credit risk information
could change according to the level of credit risk. Specically, by using BQR, we establish a hybrid credit risk model with dynamics
loadings. We judge the performance of the proposed model by comparing it with that of the conventional logit model with
uniform loadings.
There is a rapidly expanding empirical literature on quantile regression (QR) in both economics and nance. For example:
Chamberlain (1994) and Buchinsky (1994, 1997) on wage effects; Conley and Galenson (1998) and Gosling et al. (2000) dealing
with earnings inequality and mobility; Taylor (1999), Engle and Manganelli (1999), Chernozhukov and Umantsev (2001), and
Bassett and Chen (2001) on value at risk. Kordas (2006) extends the maximum score estimation method introduced by Manski
(1975, 1985) to establish the BQR model for binary data. In this study, we borrow the BQR technique in developing a dynamic
hybrid model of credit risk. It is exible enough to capture any non-uniform relation between the probability of default and the
explanatory variables. Besides, unlike conventional hybrid credit risk model, it allows for the segmentation of the data together
with the corresponding factor loadings to be endogenously determined. Specically, we implement a BQR model in which the
optimal loadings vary with the inherent default risks of the borrowers.
3. Model specications
3.1. Accounting-ratio-based and market-based information
3.1.1. Altman's z-score (accounting-ratio-based information)
To capture the accounting-ratio-based information, the widely used z-score variable derived by Altman (1968) is adopted in
this study. Specically, a rm's z-score is calculated as follows:
z = 1:2x1 + 1:4x2 + 3:3x3 + 0:6x4 + 1:0x5

x1
x2
x3
x4
x5

= working capital / total assets

= retained earnings / total assets
= earnings before interest and taxes / total assets
= market value of equity / total assets
= sales / total assets

We adopt the z-score equation exactly as it was proposed by Altman (1968). Specically, we use the same numerical values of
the weights of the nancial ratios as obtained by Altman.8 One might argue that banks adopting Altman's approach should
recalibrate the equation against the default experiences of their own credit portfolios to achieve the highest goodness-of-t.9
Given the fact that the focus of this study is to compare the importance of the accounting-ratio-based and market-based
information rather than to pursue the best credit risk model, we do not recalibrate the weights in this analysis.
8

In practice, some banks also use the z-score equation exactly as it was proposed by Altman in their risk rating systems.
For example, an optimal set of weights can be obtained by maximizing the difference between the average score of those borrowers that later default and the
average score of those that do not.
9

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

821

3.1.2. Distance-to-default (market-based information)

The distance-to-default (DD) variable derived from Merton's model is employed to capture the market-based information. The
DD of a rm can be computed by knowing both the current level and volatility of its market equity value. The calculation of DD is
detailed below.
If we assume a company's asset value (A1) at the end of the year follows a normal distribution with a mean equals to the current
asset value (A0) and a standard deviation equals to A, the probability of default (i.e., its equity value, E1, becomes equal to or less
than zero) is simply:






A D
E
PrE1 0 = 0
= 0
2
A
A
where E0 and D are respectively the current equity value and the constant level of debt; whereas () is the cumulative normal
probability function.10 The ratio E0/A is referred as the distance-to-default (DD). From Eq. (2), the higher (lower) the level of the
current equity value (the volatility of asset value), the larger the value of DD, and thus the lower the probability of default. Given
that a company's asset value is difcult to be observed directly, we use the volatility of equity value (E) as a proxy of the volatility
of asset value (A) in this study.11 Equity value is readily available for publicly traded companies, and it reects the market's
collective opinion of the prospect of its business.
We estimate E0 by computing the mean value of equity over a quarterly period. The standard deviation (E) of equity values are
also estimated using daily data observed during that quarter. The DD is therefore a function of both the mean and standard
deviation of the equity value over that quarter. At the end of each quarter, the two parameters are updated with the most recent
market information.
3.2. Hybrid model with constant loadings
Intuitively, information for both accounting-ratio-based z-score and market-based DD could be captured by conducting a
regression in which both the z-score and DD are used as explanatory variables. Similar to many other studies (e.g. that of Lfer,
2007), we rst consider the conventional logit regression model in establishing the hybrid bankruptcy prediction model with
constant loadings. In particular, we rst denote the credit risk of company i by a single number yi (i.e., a credit score), which can
be expressed as a weighted sum of a number of observable variables of the loan applicant.
0
*
yi = xi + ui
3
where xi is the a (K 1) vector of explanatory variables of yi and is an (K 1) unknown vector of parameters (i.e. loadings) to
be estimated.
Notably, yi is an unobservable latent variable. What we observe is a dummy variable yi which is dened as: yi = 1 if yi N 0 (i.e.,
company i defaults); otherwise, yi = 0 (i.e., company i does not default). The default probability for company i is thus dened
as:
0

Pi = Pyi = 1 jxi = Pxi + ui N 0 = Pui N xi

Assuming the distribution of ui is symmetric, we can write:

Pi = Fxi

where F is the cumulative distribution function of ui. Subsequently, if the cumulative distribution of ui is logistic, we have what is
known as the logit model and default probability becomes12:
Pi = pyi = 1jxi =

1
0

1 + exi

To create the best model, we want to nd the set of weights that produces the best t between Pi and the observed default
events. Specically, we would like Pi to be close to 100% for those companies that eventually default, whereas close to 0% for those
who do not. In the subsequent analysis, we estimate the optimal weights by using maximum likelihood estimation (MLE).13
To capture both the accounting-ratio-based and market-based information, this study adopts a regression model in which both
the z-score and DD variables are used as the explanatory variables for yi. We use the model specication of Eq. (3) to denote a
10

The second equality of Eq. (2) is based on the assumption that the value of assets equals to the value of debt plus that of equity.
In the literature, different formulations of the distance-to-default have been proposed and implemented. Since the objective of this study is to compare the
signicances of the accounting-ratio-based and market-based information rather than to pursue the best credit risk model, we adopt a basic formulation based
on the original Merton's model which can be implemented relatively easily. Besides, we simply consider the volatility of equity value as a direct substitute for the
volatility of asset value. One could have theoretically established the relation between the two volatilities based on certain assumptions. For example, in deriving
DD for computing a company's expected default frequency, Moody's KMV derives the volatility of asset value from the observed volatility of equity value via a
transformation utilizing information like the debt value of the company.
12
If ui follows a normal distribution, we have the probit model.
13
The MLE method is well documented in the literature, and therefore this study omits any detailed discussion of it.
11

822

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

hybrid model with constant weights, which has been commonly used in previous related research. Such a specication is
potentially restrictive owing to the use of constant loadings for the explanatory variables. For example, in adopting this model, we
assume that the sensitivities (i.e., ) of default risk to the explanatory variables are identical for all companies, disregarding
whether they are of good or poor credit quality.
3.3. Hybrid model with dynamic loadings
To resolve the potential shortcoming of the restriction of constant loadings on the z-score and DD variables, this study adopts
the BQR technique in establishing the following hybrid model with dynamic loadings.
0
*
yi = xi + ui

0
*
Quant yi jxi inf fy* : Fi y* jxg = xi

Quant ui jxi = 0

where Quant(yi |xi) denotes the th conditional quantile of yi on the regressor vector xi; is the unknown vector of parameters
to be estimated for different values of in (0,1); ui is the error term which follows a continuously differentiable cumulative
density function Fu(.|x) and a density function fu(.|x).14 The value Fi(.|x) denotes the conditional distribution of y given x. Varying
the value of from 0 to 1 reveals the entire distribution of y conditional on x.
Using the score function method proposed by Manski (1975, 1985), the estimator for is obtained from:
argmin

yi Ifxi 0g

10

i=1

where yi is the observable status of company i (yi = 1 if default; yi = 0 if not default), N is the number of observations, I{} is the
indicator function, and (v) [ I{v b 0}]v is the asymmetric absolute loss function of Koenker and Bassett (1978). We estimate
the BQR model using the simulated annealing algorithm of Corana et al. (1987) and Goffe et al. (1994). The standard errors of
estimates are obtained via bootstrapping method.15
One feature of the BQR model used in this study is the ability to trace the entire distribution of the dependent variable (y)
conditional on the independent variable (x). In particular, comparing Eq. (7) with Eq. (3) reveals a key feature of BQR technique:
the estimator vector of varies with . We can therefore examine the variation of the dynamic loading vector (), at different
quantile levels of credit risk (y). In the following discussion, the model specications of Eq. (7) are applied to a hybrid credit risk
model with dynamic loadings.
The default probability generated by the BQR model is detailed below. First, according to the denition of the binary response
model, the probability of the event {yi = 1} is presented as:
0

Pi = pyi = 1jxi = Ifxi + ui 0gdFui j xi

11

where Fui|xi is the distribution of ui conditional on xi. Subsequently, we rewrite Eq. (11) as:
1

Pi = 0 Ifxi 0gd

12

which gives the probability of success as a function of the quantile process , 2 (0, 1).
In this study, we establish the bankruptcy prediction model with dynamic loadings over nineteen quantile levels: 0.05, 0.10,
0.15, 0.20, 0.25, 0.30, 0.35, 0.40, 0.45, 0.50, 0.55, 0.60, 0.65, 0.70, 0.75, 0.80, 0.85, 0.90 and 0.95, and thus nineteen sets of loadings
are estimated. In this setup, we can simplify Eq. (12) as:
1 0:95
0
Ifx = jj jj0g
19 = 0:05 i
where ||.|| denotes the Euclidean norm and /|||| are the normalized coefcients.
Pi =

13

4. Model evaluation approaches

To compare the performances of alternative bankruptcy prediction models, we draw the cumulative accuracy proles (CAP)
and operating characteristic (ROC) curves, together with computing the respective accuracy ratios.16
14
Following Kordas (2006), the conditional errors in the BQR model are assumed to be independently and identically distributed (i.i.d.). Under i.i.d. errors,
Kordas (2006) have dened the BQR estimator and proved its consistency and asymptotic normality.
15
In Appendix A, we outline the methodology of estimating the standard errors via bootstrapping. The program codes (in version 6.1 of S-plus) used in the
computations are available (upon request) from the authors.
16
The CAP and ROC curves are commonly-used techniques for assessing the discriminatory power of risk rating methodologies.

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

823

4.1. CAP curve

To plot the CAP curve, we rst rank companies by their default probabilities (i.e., credit scores) as predicted by the model, from
highest to lowest. Then, out of those companies with a score higher than a value such that altogether they represent x% of the total
number of companies, we record the corresponding number of defaulted companies being captured as a percentage (y%) of total
number of defaulted companies. The CAP curve can then be traced out by varying x from 0 to 100 and plotting the corresponding
values of x and y along and the x-axis and y-axis, respectively. Fig. 1 shows an example of a CAP plot.
Using a good model will result in a majority of the defaulters having relatively high default probability estimates and so the
percentage of defaulters being captured (the y values in Fig. 1) increases quickly as one moves down the sorted sample of all
companies (the x values in Fig. 1). If the model were totally uninformative, for example, by assigning default probabilities
randomly, we would expect to capture a proportional fraction (i.e., x% of the defaulters with about x% of the observations),
resulting in a CAP curve along the 45-degree line (i.e., the Random CAP curve of Fig. 1). A perfect model would produce the Ideal
CAP curve, which consists of a straight line capturing 100% of the defaults within a fraction of the population equal to the default
rate of the sample.
While CAP plots are a convenient way to visualize the model performance, it will be even more convenient if we can summarize
the predictive accuracy of the model in a single statistic. One such statistic is the accuracy ratio derived from the CAP curve:
Accuracy ratio by CAP curve = the area under a models CAP = the area under the ideal CAP
The higher the accuracy ratio, the better the predictive power of the model.
4.2. ROC curve
The ROC curve is constructed by varying the cutoff probability. In particular, for every cutoff probability, the ROC curve denes
the true positive rate (percentage of defaults that the model correctly classies as defaults) on the y-axis as a function of the
corresponding false positive rate (percentage of non-defaults that are mistakenly classied as defaults) on the x-axis.17 As shown
in Fig. 2, the ROC curve of a constant or entirely random prediction model corresponds to the 45-degree line, whereas a perfect
model will have a ROC curve that goes straight up from (0, 0) to (0, 1) and then across to (1, 1). Given two models, the one with
better performance will display a ROC curve that is further to the top left of Fig. 2 than the other.
The accuracy ratio of a model obtained with the ROC curve is dened as the area under its ROC curve, i.e.:
Accuracy ratio by ROC curve = 2 area under a models ROC curve0:5
It must be noted that a model with perfect performance has an accuracy ratio of one (all defaulting rms are assigned a larger
probability of bankruptcy than any surviving rms), whereas a model with constant or random predictions has an accuracy ratio of
0. In general, models with higher accuracy ratios exhibit better performance on bankruptcy prediction.
5. Empirical results
5.1. Data
Our sample of failed (i.e., defaulted) rms consists of the rms encountering bankruptcy or liquidation events as dened by the
Compustat database over the period from 1996 Q2 to 2006 Q4.18 It consists of a total of 73 failed rms.19 To ensure our sample of
non-failed rms is of comparable size, we only consider a relatively small fraction of all the non-failed rms in Compustat
database.20 Specically, we select those non-failed rms in the Compustat database which are also Fortune 500 rms with a Brating or higher during the same time period as our sample of non-failed rms.21 This gives us a total of 138 rms in our sample of

17
For example, if the cutoff is equal to 10%, a rm of which the expected probability of default exceeds 10% is predicted to be a defaulted rm by the model,
whereas a rm of which the expected probability of default is less than 10% is predicted to be a non-defaulted rm. The true positive rate and false positive
rate are then computed by expressing the number of correctly and incorrectly predicted defaulters respectively as percentages of the total numbers of defaulted
and non-defaulted rms being realized.
18
We do not consider a very long sample period because we are assuming the loadings of z-score and DD are constant throughout the sample period. This
assumption will become less valid if we consider a sample which spans over an extended time period. We do not want our ndings to be affected by the potential
violation of this assumption of time consistency.
19
The list of all 73 failed companies selected for this study is given in Appendix B.
20
Some previous studies (e.g. Altman, 1968) also consider balanced samples of failed and non-failed rms. Using samples of comparable sizes in this study
ensures that similar weights are assigned to these two types of rms in examining the explanatory powers of z-score and DD. It therefore enhances our ability in
detecting any difference in the predictive power of alternative default risk models.
21
Using rms listed in the Fortune 500, which are arguably of relatively better credit quality, as our non-failed rms ensures our samples of failed and nonfailed rms are very different from each other in terms of their credit qualities. Fortune 500 rms with B-rating or higher are deemed to be of good nancial
health and far from experiencing any distressed conditions. The discriminatory power of the models examined in this study could therefore be enhanced, thus
enabling us to achieve a more rened comparison of the performances of alternative models. Focusing on Fortune 500 rms also ensures the qualities of the stock
price and nancial statement information used in computing z-score and DD in this study.

824

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

Fig. 1. Illustrations of CAP curves.

non-failed rms. Financial rms are excluded from both samples. During the sample period, we collected a total of 1329 and 4959
observations of quarterly nancial data of the above samples of failed and non-failed rms respectively. Equity values and nancial
statement information are obtained from the CRSP and Compustat databases. Table 1 presents the average values of the z-score
and DD. Comparing the values of failed rms against non-failed rms, the z-score and DD values of the latter are considerably
higher than those of the former.

5.2. Model estimation results: Constant loadings versus dynamic loadings

Table 2 presents the estimation results of the hybrid model with constant loadings derived by the logit regression model. Using
the 5% signicance level as a criterion, the coefcients of both the z-score and DD are signicantly negative.22 This result indicates
that as a company's z-score and DD increases, it is less likely to go bankrupt. Although the two estimates are signicant and have
the expected sign, this hybrid model derived by the conventional logit regression approach does not allow for the impact of z-score
and DD variables to differ between good and poor quality companies.
In Table 3, we report the estimation results of the hybrid model with dynamic loadings derived by the BQR model. The negative
values of the point estimates together with their respective 95% condence intervals are plotted in Figs. 3 and 4. This model allows
for the loadings of the z-score and DD to assume different values based on the quantile levels of credit risk. Specically, we
estimate the loadings at nineteen different quantile levels: 0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35, 0.40, 0.45, 0.50, 0.55, 0.60, 0.65,
0.70, 0.75, 0.80, 0.85, 0.90 and 0.95.
First, let us examine the conditional quantile estimates for the loadings of the z-score in credit risk prediction. As expected, the
z-score variable has a negative loading at each quantile level being considered. The coefcients are however of different magnitude
and degree of statistical signicance. In particular, except for the 0.25 quantile, the loadings on the z-score are found to be
statistically signicant at the 5% level for all of the low quantile levels below or equal to 0.60. The loadings however become
insignicant at higher quantile levels from 0.65 to 0.95.
Moreover, as shown in Fig. 3, the magnitudes of the z-score loading estimates change dramatically across the quantiles.23 In
particular, when moving up the quantile levels, the magnitude decreases considerably and eventually approaches zero at the
quantile levels of 0.85, 0.90 and 0.95. These results suggest that the accounting-ratio-based information, such as the z-score, is
more useful in predicting bankruptcy of those companies of relatively better credit quality (i.e., low quantile level). The ability of
the z-score in predicting bankruptcy of poor quality companies (i.e., high quantile level) is however neither statistically nor
economically signicant.24
The above ndings require some intuitive elucidation. First, accounting fraud is potentially involved in poor quality
companies and hence their z-scores (computed based on their nancial statements) are more likely to be unreliable. Second,
we actually do not expect the sensitivity of default probability to accounting ratios is uniform across companies of different
22

One-period lag z-score and DD variables are employed as explanatory variables for rm bankruptcy prediction.
As expected, the point estimates of the loadings of z-score and DD variables are negative at each quantile level being considered. In Fig. 3, for illustrative
purpose, we present the negative of these negative values. That is, Fig. 3 presents the magnitudes of the point estimates of these loadings.
24
The estimated loadings of the z-score variable become insignicant in the higher quantiles from 0.65 to 0.95 (i.e., their 95% condence intervals overlap the
value of zero in Fig. 3).
23

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

825

Fig. 2. Illustrations of ROC curves.

credit qualities.25 Let us take EBIT/Assets ratio (which is one of the ve accounting ratios used in formulating the z-score) as an
example. Theoretically, it is negatively related to the company's default probability. Whereas Company X having a EBIT/Assets
ratio of 5% could have a much lower probability of default than Company Y having a 5% EBIT/Assets ratio, the difference
between the probability of default of the above Company Y and another Company Z having a ratio of 15% might not be as
signicant. Given their negative EBIT/Assets ratios, the default risks of both Company Y and Z are likely to be high, thus
belonging to the high quantiles of the BQR regression. The fact that Company Y has a less negative EBIT/Assets ratio than
Company Z does not necessarily suggest that it is in a much better nancial position to avoid a potential bankruptcy event.
Consequently, the negative relation between EBIT/Assets and default probability could be disappearing for the high quantile
levels (as illustrated in Fig. 3).
Next, the coefcients of the DD variables, as shown in Table 3, are signicantly negative at the 5% level for most of the higher
quantile levels (the loading of the 0.80 quantile is statistically signicant at 10% level); whereas they become insignicant at the
three lower quantiles, namely 0.05, 0.10 and 0.25 (i.e., as shown in Fig. 4, their 95% condence intervals overlap the value of zero).
The magnitudes of the DD loadings tend to be increasing with the quantile levels. The average magnitude of the DD loadings of the
lowest 6 quantiles is 0.12; whereas that of the highest 6 quantiles being 0.16. This increasing trend of the magnitude of the DD
loadings is consistent with the implications of the Merton's model. In particular, under the assumption of normally distributed
asset value, the probability of default is equal to (DD), as presented in Eq. (2). Given this functional relation, the magnitude of
the rate of change of probability of default with respect to DD increases as DD decreases. We therefore expect the magnitude of the
DD loading increases with the quantile level. In Fig. 5, we plot the negative values of the rates of changes of probabilities of default
implied by Merton's model over the range of probabilities of default from 2% to 20%. The magnitudes of the point estimates of the
DD loadings, which are also plotted on Fig. 5, tend to vary with the quantile levels in a similar increasing trend. The constant
loading model as depicted in Eq. (3) is likely to be too restrictive in capturing this non-uniform relation between DD and the
probability of default.
To summarize, two hybrid models using both the z-score and DD variables have been examined. In the rst model (dened by
Eq. (3)), we impose constant weights on the two variables and thus disregard the possibility that credit risk prediction might be
more reliable with market-based variables than with accounting-ratio-based information in certain conditions, and vice versa. In
the second hybrid model (dened by Eq. (7)), we adopt the BQR technique to allow the loadings of the z-score and DD variables to
differ among companies of different credit risks. Our empirical ndings suggest that creditors should decrease (increase) the
loading of the accounting-ratio-based z-score while increase (decrease) the loading of the market-based DD when they appraise
companies which are perceived to be of higher (lower) credit risk.26

25

The relation between default probability and accounting ratios might not even be monotonic.
Our ndings suggest that the magnitude of the loading |z,| (|DD,|) decreases (increases) when the quantile level increases. However, the products
|z,| z-score and |DD,| DD should both be decreasing when we move up the quantile level. Note that, the lower the values of the z-score and DD, the higher
is the default risk of the rm (i.e. belonging to a higher quantile). Our ndings therefore suggest the diminishing effect of DD as a default risk indicator when DD
decreases in value is partially offset by an increasing value of |DD,| as we move up the quantile level. Our BQR approach enriches the modeling of default risk by
allowing for the capturing of this non-linear effect which cannot be modeled under a constant loading model. The authors acknowledge the anonymous reviewer
for this comment.
26

826

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

Table 1
Average values of z-score and DD: failed rms versus non-failed rms.

Altman's z-score
DD (distance-to-default)

Failed rms

Non-failed rms

9.46
3.08

7.25
22.90

Failed rms are those which experienced bankruptcy/liquidation events as dened by the Compustat database over the period from 1996 Q2 to 2006 Q4. Firms
listed in the Fortune 500 with at least a B-rating during the same time period are classied as non-failed. Financial rms are excluded from the sample. The overall
sample consists of a total of 74 failed and 138 non-failed rms. During the sample period, we collected a total of 1329 and 4959 observations of quarterly nancial
data of the above samples of failed and non-failed rms respectively.
The average values of the z-score and DD of non-failed rms are considerably higher than those of failed rms.

5.3. Bankruptcy prediction tests of alternative risk models

The above empirical results have showed how the loadings of accounting-ratio-based and market-based information change
according to the credit quality of companies. Specically, our results indicate that creditors should pay more attention to the
market-based information when assessing a rm with higher credit risk. In contrast, for a rm with lower credit risk, the
accounting-ratio-based z-score deserves more attention as a bankruptcy prediction variable. The loading of each bankruptcy
forecasting variable varies according to the credit risk level of the company.
The remaining inquiry is whether the hybrid model with dynamic loadings offers banks better bankruptcy prediction
performance. To compare the bankruptcy forecasting performances of the two models, we conduct two commonly used tests: CAP
and ROC curves, and the corresponding accuracy ratios.
Fig. 6 shows the CAP curves for the two bankruptcy prediction models established in this study. The 45 degree line represents
the nave case (which is equivalent to a random assignment of scores). Apparently, the two models perform better than a random
model; however, the relative performance of the two competing models is not clear. In particular, the model with dynamic
loadings derived by the BQR model performs better than the logit model with constant loadings in the bottom 50% of the
population in terms of credit quality. Notably, the model with dynamic loadings performs slightly worse than the model with
constant loadings in the case of good quality rms (that is, the top 50% of the population).
In Fig. 7, we present the ROC curves for the two competing models. It shows two things: (i) the two credit models selected in
this study are considerably better than the nave model, and (ii) the dynamic BQR model outperforms the logit model with
constant loadings, resulting in a larger area under the ROC curve.
In Table 4, we summarize the results of the accuracy ratio (AR) as a performance measure of bankruptcy prediction. The model
with dynamic loadings derived by the BQR technique is consistently associated with higher AR values in comparison with the
setting with constant loadings generated by the traditional logit model. The conclusion is clear: adopting a hybrid model with
dynamic loadings does enhance bankruptcy prediction performance.
5.4. Out-of-sample tests
The above results demonstrate that the model with dynamic loadings would provide better prediction in rm bankruptcy in an
in-sample analysis. Specically, the tests are conducted on the same historical sample of default data on which the two models
were calibrated against. It is not surprising that, by providing more degree of freedom, we achieve a better in-sample goodness-oft when adopting the dynamic BQR model. Undeniably, banks would be more concerned with how well the alternative forecasting
techniques will perform in the future (i.e., in an out-of-sample setting). For that purpose, we conduct an out-of-sample test by
randomly withholding 500 observations (150/350 of them are failed/non-failed rms), which are dened as the test set. The
residual observations are dened as the model set and are used in calibrating the models and calculating the loadings. The
calibrated models are then used to predict the default events in the previously withheld test set.

Table 2
Estimates of bankruptcy prediction model with constant loadings derived by logit regression model.
Variables

Expected sign

Estimate (standard error)

Intercept
z-score
DD

ve
ve

1.9576 (0.0856)*
0.0313 (0.0055)*
0.2560 (0.0077)*

This table presents estimation results of the hybrid bankruptcy prediction model with constant loadings derived by the logit model (see Eqs. (3)(6) for the details
of model specications).
The value in the parenthesis is the standard error of the point estimate and the * denotes statistical signicance at 5% level. As shown in this table, the coefcients of
z-score and DD are signicantly negative. This result indicates that as a company's z-score and DD increases, it is less likely to go bankrupt. Although the two
estimates are signicant and have the expected sign, this credit risk model derived by the conventional logit regression approach does not allow for the loadings of
z-score and DD variables to differ between good and poor quality companies.

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

827

Table 3
Estimates of bankruptcy prediction model with dynamic loadings derived by BQR model.
Quantile

Intercept

z-score

0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.45
0.50
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0.90
0.95

0.8821
1.1032
1.7920
2.677
1.1110
2.4856
1.8971
2.6936
2.2890
2.1434
2.4698
2.5100
1.9440
2.7941
1.3532
1.0193
1.4426
2.9101
2.7336

1.4214
1.282
1.6957
1.763
0.7124
1.5610
1.1482
1.5920
1.0936
1.0463
1.2055
1.2249
0.2848
0.0797
0.0413
0.0309
0.0054
0.0075
0.0097

(0.5523)
(0.5195)*
(0.5783)*
(0.5374)*
(0.6821)
(0.4448)*
(0.4520)*
(0.2992)*
(0.4349)*
(0.5775)*
(0.4435)*
(0.4872)*
(0.6554)*
(0.4582)*
(0.6277)
(0.8050)
(0.4839)*
(0.3801)*
(0.5119)*

DD
(0.7077)*
(0.5849)*
(0.5446)*
(0.4607)*
(0.4865)
(0.2519)*
(0.2875)*
(0.1965)*
(0.2095)*
(0.2971)*
(0.2859)*
(0.3300)*
(0.4183)
(0.2531)
(0.0915)
(0.0355)
(0.0547)
(0.0056)
(0.0058)

0.0519
0.0725
0.1309
0.2255
0.0801
0.1807
0.1415
0.1934
0.1782
0.1347
0.1552
0.1577
0.1674
0.2778
0.1201
0.0905
0.1165
0.1948
0.1417

(0.0365)
(0.0454)
(0.0536)*
(0.0505)*
(0.0554)
(0.0381)*
(0.0368)*
(0.0245)*
(0.0379)*
(0.0431)*
(0.0292)*
(0.0434)*
(0.0500)*
(0.0604)*
(0.0625)*
(0.0545)
(0.0430)*
(0.0270)*
(0.0289)*

This table presents the estimation results of the hybrid bankruptcy prediction model with dynamic loadings derived by BQR model (see Eqs. (7)(13) for the
details of the model specications).
In this model, the loadings of the z-score and DD are allowed to change according to the quantile level of credit risk. In particular, the loadings are estimated at
nineteen different quantile levels: 0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35, 0.40, 0.45, 0.50, 0.55, 0.60, 0.65, 0.70, 0.75, 0.85, 0.90 and 0.95.
The value in the parenthesis is the standard error of the point estimate and the * denotes statistical signicance at the 5% level.
The z-score variable has a statistically signicant loading at the lower quantile levels, from 0.05 to 0.60 (except for the 0.25 quantile); it becomes insignicant at
higher quantile levels from 0.65 to 0.95.
The coefcients of the DD variable are signicantly negative at the 5% level for most of the higher quantile levels (the loading of the 0.80 quantile is statistically
signicant at 10% level); nevertheless, they become insignicant at the three lower quantiles, such as 0.05, 0.10 and 0.25.

The estimates of bankruptcy prediction model with dynamic loadings derived by BQR model using the 5788 (=6288 500)
residual observations from the model set are listed at Table 5. Importantly, using the 5% signicance level as a criterion, the zscore variable is insignicant at the six higher quantile levels, from 0.70 to 0.95 whereas it is signicant under the lower quantile
levels, from 0.05 to 0.65 with two exceptions: 0.30 and 0.50 quantiles. By contrast, although the DD variable is signicant for most
quantile levels, it becomes insignicant at the two lower quantiles: 0.05 and 0.30. Our previous nding of the loadings of z-score
(DD) variables becoming less signicant for companies with worse (better) credit quality is therefore robust in this subsample of
the dataset. Table 6 presents the results of the out-of-sample bankruptcy prediction performances of the two models. First, in

Fig. 3. The negative value of z-score loading estimate and its condence intervals across various quantile levels.

828

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

Fig. 4. The negative value of DD loading estimate and its condence intervals across various quantile levels.

Fig. 5. The negative values of DD loadings implied by Merton's model versus the DD loading estimates by the BQR model.

having positive AR values, the two hybrid models still outperform a nave (random) model. Second, by having higher AR values,
the dynamic BQR model still outperforms the logit model with constant loadings.27
6. Conclusions and future research
While adopting the BQR technique, we propose a bankruptcy prediction model with dynamic loadings for both the accountingratio-based z-score and the market-based DD variable. With this proposed model, we conduct an empirical study on a dataset of
default events observed during the period from 1996 to 2006. Our sample of defaulted rms consists of those rms which

27
We also re-run the tests using the probit model (to conserve space, we do not present these results here). The AR results are qualitatively similar when the
probit model is used instead of the logit model. This result indicates that bankruptcy prediction enhancement heavily depends on allowing for dynamic loadings
on accounting-ratio-based and market-based information, rather than on the distribution assumption of the errors ui in Eq. (3).

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

829

Fig. 6. CAP curves for the two hybrid models.

Fig. 7. ROC curves for the two hybrid models.

experienced bankruptcy or liquidation events as recorded in the Compustat database. Our sample of survived rms is made up of
the Fortune 500 companies with over a B-rating over the same time period.
The empirical results of this study are consistent with the following notions. First, in predicting the bankruptcy of those
companies with relatively poor (good) credit quality, we can improve the accuracy by putting more (less) emphasis on the
market-based DD variable while reducing (increasing) the emphasis on the accounting-ratio-based z-score. Second, the proposed

Table 4
Accuracy ratios comparison: in-sample tests.
The setting with constant loadings derived by the Logit Model The setting with dynamic loadings derived by the BQR Model
Accuracy ratio by CAP curve 80.38%
Accuracy ratio by ROC curve 79.47%

85.72%
84.60%

This table summarizes the results of accuracy ratio (AR) as performance measure of bankruptcy prediction. See Section 4 in the text for detail discussions on the
two AR measures.
The AR of a nave (random) model is zero.

830

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

Table 5
Estimates of bankruptcy prediction model with dynamic loadings derived by BQR model using observations from the model set.
Quantile

Intercept

z-score

0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.45
0.50
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0.90
0.95

1.5778
2.0386
2.7618
2.1054
2.7445
1.0549
1.5218
2.8376
2.6112
0.7964
2.3000
2.9256
2.2818
1.6895
2.7199
1.9171
1.5979
2.8552
1.8947

2.5297
1.8370
2.4807
1.3898
1.6605
0.6116
0.9080
1.6926
1.4354
0.3888
1.1244
0.9433
0.7371
0.1934
0.0821
0.0587
0.0060
0.0173
0.0051

(0.4936)*
(0.4579)*
(0.5559)*
(0.3666)*
(0.5666)*
(0.7683)
(0.5919)*
(0.4342)*
(0.5058)*
(0.8296)*
(0.5029)*
(0.5002)*
(0.5947)*
(0.4000)*
(0.4675)*
(0.4545)*
(0.4680)*
(0.5490)*
(0.5362)*

DD
(0.6080)*
(0.5670)*
(0.5435)*
(0.3189)*
(0.3508)*
(0.4666)
(0.3237)*
(0.2977)*
(0.3415)*
(0.4564)
(0.3169)*
(0.3826)*
(0.4101)*
(0.1193)
(0.0912)
(0.0591)
(0.0607)
(0.0120)
(0.0066)

0.0942
0.1531
0.2085
0.1773
0.2048
0.0805
0.1087
0.2027
0.1889
0.0500
0.1444
0.1941
0.1512
0.1345
0.2416
0.1700
0.1290
0.1839
0.0950

(0.0604)
(0.0393)*
(0.0420)*
(0.0310)*
(0.0497)*
(0.0606)
(0.0465)*
(0.0342)*
(0.0418)*
(0.0669)*
(0.0429)*
(0.0345)*
(0.0368)*
(0.0421)*
(0.0475)*
(0.0408)*
(0.0336)*
(0.0354)*
(0.0265)*

This table presents the estimation results of the hybrid bankruptcy prediction model with dynamic loadings derived by BQR model (see Eqs. (7)(13) for the
details of the model specications).
To conduct the out-of-sample test, 500 observations (127/373 of them are failed/non-failed rms) are randomly withheld and dened as the test set. The
residual 5788 (= 6288 500) observations are dened as the model set and are used in calibrating the models and calculating the loadings.
Other notions are consistent with Table 3.

Table 6
Accuracy ratios comparison: out-of-sample tests.
The setting with constant loadings derived by the Logit Model The setting with dynamic loadings derived by the BQR Model
Accuracy ratio by CAP curve 81.75%
Accuracy ratio by ROC curve 81.69%

88.00%
85.72%

To conduct the out-of-sample test, 500 observations (127/373 of them are failed/non-failed rms) are randomly withheld and dened as the test set. The
residual observations are dened as the model set and are used in calibrating the models and calculating the loadings. We then test the performances of the
calibrated models in predicting the default events within the test set.
The AR of a nave (random) model is zero.

model performs better than a hybrid model with constant loadings in predicting default events in both the in-sample and out-ofsample settings. This study provides both the theoretical and empirical underpinnings of a dynamic hybrid model which is more
able to explain and predict the default events of companies of diverse credit qualities than conventional logit model. We therefore
provide an alternative modeling approach for banks to consider in developing their internal risk rating systems.
One important caveat should be noted in interpreting the results of this study. In practice, default dependencies among rms
play an important role in the quantication of a portfolio's credit risk exposure. Moreover, growing linkages in nancial markets
also have led to a greater degree of joint default propensity. Some of the systematic changes in credit risk, which govern the
probability of the occurrence of multiple default events, might not be able to be fully captured by either the z-score or the DD
variable. The conditional distribution of the error term in the BQR model therefore might not be i.i.d. Dealing with default
dependencies across rms and over times by relaxing the assumption of i.i.d. errors is a valuable direction for future research.
Acknowledgment
The authors gratefully acknowledge funding from the National Science Council of Taiwan (NSC96-2416-H-006-023-MY3).
Appendix A. The bootstrap estimate of the standard error
Assume we have a real-valued estimator (X1, X2,, Xn), which is a function of n independently and identically distributed
observations:
iid

X1 ; X2 ; :::; Xn F;

A1

F being an unknown probability distribution on a space . Having observed X1 = x1, X2 = x2,, Xn = xn, we wish to obtain an
estimate of the standard error of .

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

831

The true standard error of is a function of F, n, and the form of the estimator , say
:; :; :::; : = F:
F; n;

A2

This last notation emphasizes that, knowing n and the form of , the true standard error is only a function of the unknown
distribution F.
The bootstrap estimate of the standard error, B, is simply
;
= F

A3

where F is the empirical probability distribution

1
F : mass on xi ;
n

i = 1; 2; ; n:

A4

In practice, the function (F) is usually impossible to express in simple form, and B must be evaluated using a Monte Carlo
algorithm:
Step 1. Construct F as at Eq. (A4).
Step 2. Draw a bootstrap sample from F ,
iid
X1 ; X2 ; :::; Xn F;

A5

= (X1, X2,..., Xn).

and calculate
Step 3. Independently repeat Step 2 some number B times, obtaining bootstrap replications (1), (2),, (B), and
calculate
"
#1 = 2

B
*
b *

B =

;
A6
B1
b=1
where
B

b = B:
= *
*

A7

b=1

As B , the right-hand side of Eq. (A6) converges to (F ).

Appendix B. Alphabetical list of the failed companies selected for this study
No.

Firm name

1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.

A3DO Co
Acclaim Entertainment Inc
Accrue Software Inc
All American Semiconductor
Allied Holdings Inc
Aphton Corp
Armstrong Holdings Inc
Biotransplant Inc
Boundless Corp
Calpine Corp
Ceyoniq AG
Cinemastar Luxury Theaters
Collins & Aikman Corp
Composite Technology Corp
Congoleum Corp
Corporacion Durango SA
D G Jewelry Inc
Dairy Mart Convenience Strs
Dana Corp
Delphi Corp
Donnkenny Inc
DT Industries Inc
Dura Automotive Sys
Earthshell Corp
(continued on next page)

832

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

Appndix B (continued)
No.

Firm name

25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.
65.
66.
67.
68.
69.
70.
71.
72.
73.

Enernorth Industries Inc

Enesco Group Inc
Engage Inc
Environmental Elements Corp
Fantom Technologies Inc
Federal Mogul Corp
Florsheim Group Inc
Frisby Technologies Inc
Frontline Capital Group
Global Power Equipment Group
Glycogenesys Inc
Grace W R & Co
Hancock Fabrics Inc
Home Products Intl Inc
Interstate Bakeries CP
Intl Fibercom Inc
KPNQWest NV
Movie Gallery Inc
National Steel Corp
Newpower Holdings Inc
NQL Inc
NX Networks Inc
Onetravel Holdings Inc
PCD Inc
Polymer Research Corp of AM
President Casinos Inc
Ramp Corp
Robotic Vision Systems Inc
Rowe Companies
Sea Containers Ltd
Solutia Inc
Storage Engine Inc
Talisman Enterprises Inc
Televideo Inc
Thermoview Industries Inc
Three Five Systems Inc
Tower Automotive Inc
Trans Industries Inc
Trinsic Inc
Tweeter Home Entmt Group Inc
Uniroyal Technology Corp
Universal Access Global Hldg
Universal Automotive Inds
V One Corp
Verilink Corp
Waterlink Inc
Westpoint Stevens Inc
Women First Healthcare
World Health Alternatives

References
Agarwal, V., Tafer, R., 2008. Comparing the performance of market-based and accounting-based bankruptcy prediction models. Journal of Banking and Finance 32
(8), 15411551.
Altman, E.I., 1968. Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. J. Finance 23, 589609.
Basel Committee, 2006. International Convergence of Capital Measurements and Capital Standards: A Revised Framework (Comprehensive Version), BIS.
Bassett, G., Chen, H.L., 2001. Quantile style: return-based attribution using regression quantiles. Empirical Econ. 26, 293305.
Black, F., Scholes, M., 1973. The pricing of options and corporate liabilities. J. Polit. Econ. 7, 637654.
Blochlinger, A., Leippold, M., 2006. Economic benet of powerful credit scoring. J. Bank. Finance 30, 851873.
Buchinsky, M., 1994. Changes in U.S. wage structure 19631987: an application of quantile regression. Econometrica 62, 405458.
Buchinsky, M., 1997. The dynamics of changes in the female wage distribution in the USA: a quantile regression approach. J. Appl. Econometrics 13, 130.
Campbell, J.Y., Hilscher, J., Szilagyi, J., 2006. In search of distress risk. Working paper, Harvard University.
Chamberlain, G., 1994. Quantile regression, censoring and the structure of wages. Adv. Econometrics 171209.
Chernozhukov, V., Umantsev, L., 2001. Conditional value-at-risk: aspects of modeling and estimation. Empirical Econ. 26, 271292.
Conley, T., Galenson, D., 1998. Nativity and wealth in mid-nineteenth-century cities. J. Econ. Hist. 58, 468493.
Corana, A., Marchesi, M., Martini, C., Ridella, S., 1987. Minimizing multimodal functions of continuous variables with the simulated annealing algorithm. ACM
Trans. Math. Softw. 13, 262280.
Engle, R., Manganelli, S., 1999. CaViaR: conditional autoregressive value at risk by regression quantiles. University of California, San Diego, Department of
Economics, Working Paper 99/20.

M.-Y.L. Li, P. Miu / Journal of Empirical Finance 17 (2010) 818833

833

Goffe, W., Ferrier, G., Rogers, J., 1994. Global optimization of statistical functions with simulated annealing. J. Econometrics 60, 65101.
Gosling, A., Machin, S., Meghir, C., 2000. The changing distribution of male wages in the UK. Rev. Econ. Stud. 67, 635666.
Hillegeist, S., Keating, E., Cram, D., Lundstedt, K., 2004. Assessing the probability of bankruptcy. Rev. Acc. Stud. 9, 534.
Kealhofer, S., 2003. Quantifying credit risk I: default prediction. Financ. Anal. J. 59, 3044.
Kealhofer, S., Kurbat, M., 2001. The default prediction power of the Merton approach, relative to debt ratings and accounting variables. Moody's KMV.
Koenker, R., Bassett, G.B., 1978. Regression quantiles. Econometrica 46, 3350.
Kordas, G., 2006. Smoothed binary regression quantiles. J. Appl. Econometrics 21, 387407.
Lfer, G., 2007. The complementary nature of ratings and market-based measures of default risk. J. Fixed Income 17, 3847.
Manski, C., 1975. Maximum score estimation of the stochastic utility model of choice. J. Econometrics 3, 205228.
Manski, C., 1985. Semiparametric analysis of discrete response: asymptotic properties of the maximum score estimator. J. Econometrics 27, 313334.
Mensah, Y.M., 1984. An examination of the stationarity of multivariate bankruptcy prediction models: a methodological study. J. Acc. Res. 22, 380395.
Merton, R.C., 1974. On the pricing of corporate debt: the risk structure of interest rates. J. Finance 29, 449470.
Miller, R., 1998. Rening ratings. Risk 9799.
Mitchell, J., Roy, P.V., 2008. Failure prediction models: performance, disagreements, and internal rating systems. European Financial Management Annual
Conference (Athens, Greece).
Oderda, G., Dacorogna, M., Jung, T., 2003. Credit risk models: do they deliver their promises? A quantitative assessment. Rev. Bank. Finance Monetary Econ. 32,
177195.
Reisz, A., Perlich, C., 2004. A market-based framework for bankruptcy prediction. Working paper, Baruch College, City University of New York.
Saunders, A., Allen, L., 2002. Credit Risk Measurement: New Approaches to Value at Risk and Other Paradigms, second ed. Wiley Finance, New York.
Stein, R., 2005. The relationship between default prediction and lending prots: integrating ROC analysis and loan pricing. J. Bank. Finance 29, 12131236.
Taylor, J., 1999. A quantile regression approach to estimating the distribution of multiperiod returns. J. Deriv. 7, 6478.
Vassalou, M., Xing, Y., 2004. Default risk in equity returns. J. Finance 59, 831868.