Financial Distress Prediction in an Interna- tional Context: A Review and Empirical Analysis of Altman’s Z-Score Model doi:10.1111/jifm.12053
Edward I. Altman NYU Salomon Center, Henry Kaufman Management Center, New York University, Stern School of Business, 44 West Fourth Street, New York, NY 10012, USA
Małgorzata Iwanicz-Drozdowska Institute of Finance, Warsaw School of Economics, al. Niepodleglosci 162, 02-513 Warsaw, Poland e-mail: [email protected]
Erkki K. Laitinen University of Vaasa, P.O.Box 700, FI-65101 Vaasa, Finland e-mail: [email protected]
Arto Suvas University of Vaasa, P.O.Box 700, FI-65101 Vaasa, Finland e-mail: [email protected]
Abstract
This paper assesses the classification performance of the Z-Score model in predicting bankruptcy and other types of firm distress, with the goal of examining the model’s usefulness for all parties, especially banks that operate internationally and need to assess the failure risk of firms. We analyze the performance of the Z-Score model for firms from 31 European and three non-European countries using different modifications of the original model. This study is the first to offer such a comprehensive international analysis. Except for the United States and China, the firms in the sample are primarily private, and include non-financial companies across all industrial sectors. We use the original Z00-Score model developed by Altman, Corporate Financial Distress: A Com- plete Guide to Predicting, Avoiding, and Dealing with Bankruptcy (1983) for private and public manufacturing and non-manufacturing firms. While there is some evidence that Z-Score models of bankruptcy prediction have been outperformed by competing mar- ket-based or hazard models, in other studies, Z-Score models perform very well. With-
The authors are grateful to the Editor, Richard Levich, and to the anonymous referees for many helpful comments and suggestions. We also wish to thank participants at the 2014 Inter- national Risk Management Conference (IRMC) in Warsaw for useful discussions. Laitinen and Suvas thank the Foundation for Economic Education, and Jenny and Antti Wihuri Foundation for financial support. Laitinen also thanks OP-Pohjola Group Research Foundation for support.
Journal of International Financial Management & Accounting 28:2 2017
© 2016 John Wiley & Sons Ltd.
out a comprehensive international comparison, however, the results of competing mod- els are difficult to generalize. This study offers evidence that the general Z-Score model works reasonably well for most countries (the prediction accuracy is approximately 0.75) and classification accuracy can be improved further (above 0.90) by using coun- try-specific estimation that incorporates additional variables.
1. Introduction
The first multivariate bankruptcy prediction model was developed by
Altman (1968) in the late 1960s. After this pioneering work, the
multivariate approach to failure prediction spread worldwide among
researchers in finance, banking, and credit risk. Failure prediction mod-
els are important tools for bankers, investors, asset managers, rating
agencies, and even distressed firms themselves. The banking industry, as
the main provider of financing in the economy, is especially interested in
minimizing the level of non-performing loans in order to maximize profit
on credit activity, and banks seek to reduce their own risk of default.
Another issue of interest for bankers is capital adequacy and the internal
ratings-based approach encouraged by the Basel Accords. The Z-Score
model has become a prototype for many of these models. Asset man-
agers and investors need reliable tools that can help them select appro-
priate companies for their portfolios. Financial distress is detrimental to
investor returns, but risk may provide opportunities for high returns on
short-sale strategies. Rating agencies assess the risk of the entities and of
securities issues, and thus, they need a tool to predict default. Altman
(1983) suggested that the management of distressed firms can utilize the
Z-Score model as a guide to financial turnaround.
The approach used for bankruptcy prediction has evolved over time.
Beaver (1966, 1968) used univariate analysis for selected ratios and
found that some had very good predictive power. Altman (1968) made
strides by developing a multiple discriminant analysis model (MDA)
called the Z-Score model. The next two decades saw additional contri-
butions to financial distress research. 1 For example, Ohlson (1980) pro-
posed a logit model, 2 Taffler (1984) offered a Z-Score model for the
United Kingdom, and Zmijewski (1984) 3 used a probit approach. Dim-
itras et al. (1996) reviewed 47 studies on business prediction models,
summarizing the methods employed and the variety of ratios used.
Discriminant analysis was the prevailing method, and the most impor-
tant financial ratios came from the solvency category, with profitability
ratios also being important.
132 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
© 2016 John Wiley & Sons Ltd
Balcaen and Ooghe (2006) reviewed 43 models of business failure pre-
diction which they classified into four categories: univariate models (1);
risk index models (2); MDA models (21); and conditional probability
models (19). However, their review omitted the rapidly growing type of
models based on option pricing theory and contingent claims (e.g., Vas-
salou and Xing, 2004; commercialized into Kealhofer, McQuown and
Vasicek’s model, known as the KMV model), as well as hazard models
(e.g., Shumway, 2001). Kumar and Ravi (2007) reviewed 128 statistical
and artificial intelligence models for bank and firm bankruptcy predic-
tions, paying special attention to the techniques used in the different
models. These authors noted that neural networks were the most popu-
lar intelligence technique. In their review, Jackson and Wood (2013) pre-
sented the frequency of occurrence of specific forecasting techniques in
the prior literature. The five most popular techniques were as follows: (1)
multiple discriminant analysis, (2) logit models, (3) neural networks, (4)
contingent claims, and (5) univariate analysis.
Recent reviews of the efficacy of these models have been offered by
Agarwal and Taffler (2008), Das et al. (2009), and Bauer and Agarwal
(2014). These reviews take into account the performance of account-
ing-based, market-based, and hazard models. These three model types
prevail in the literature. According to Agarwal and Taffler (2008),
there is little difference in the predictive accuracy of accounting-based
and market-based models; however, the use of accounting-based mod-
els allows for a higher level of risk-adjusted return on credit activity.
Das et al. (2009) showed that accounting-based models perform com-
parably to the Merton structural, market-based approach for credit
default spread (CDS) estimation. However, a comprehensive model,
which used both sources of variables, outperformed the other models.
In Bauer and Agarwal (2014), hazard models using accounting and
market information (Shumway, 2001; Campbell et al., 2008) were com-
pared with two other approaches: the original Taffler (1984) Z-score
model, which was tested by Agarwal and Taffler (2008), and a contin-
gent claims model using Bharath and Shumway’s (2008) approach.
Using U.K. data, the hazard models were superior in bankruptcy pre-
diction accuracy, ROC (Receiver Operating Characteristic) analysis,
and information content.
Even though the Z-Score model was developed more than 45 years
ago and many alternative failure prediction models exist, the Z-Score
model continues to be used worldwide as a main or supporting tool
for bankruptcy or financial distress prediction and analysis both in
Financial Distress Prediction 133
© 2016 John Wiley & Sons Ltd
research and in practice. We focus on accounting-based versions of the
Z-Score models, which even though they are occasionally outper-
formed by other models, do not rely on market data. Most firms oper-
ating in business are privately held; hence, only accounting data and
no market data (e.g., stock prices) are available. Private firms are usu-
ally financed by banks, which are obligated to assess their creditwor-
thiness and monitor their performance. In the case of internationally
active banks, from a regulatory perspective it is especially important to
use a single model for distress prediction, provisioning, and economic
capital calculation. According to current Basel regulatory require-
ments, banks need to validate their distress prediction models and doc-
ument their efficacy. Thus, it is important to analyze the performance
of accounting-based models in an international context.
In our study, we use a large international sample of firms to assess
the classification performance of the Z-Score model in bankruptcy pre-
diction. 4 We analyze the model’s performance for firms from 31 Euro-
pean and three non-European countries (China, Colombia and the
United States). These firms are mostly privately held, and a large num-
ber are from non-manufacturing industries. We use the version of the
model developed by Altman (1983) for private and public manufactur-
ing and non-manufacturing firms (the Z”-Score model). Such an exten-
sive international analysis of the Z-Score model’s performance has not
been presented to date. We regard our review and analysis as impor-
tant contributions to the economic literature.
The remainder of the paper is structured as follows. In the next sec-
tion, we summarize the original Z-Score model (Altman, 1968) and its
extension for private firms, that is, the Z’-Score and Z”-Score models
(Altman, 1983). In the third section, we present the results and conclu-
sions from the literature review on these models. The fourth section pre-
sents seven hypotheses on the performance of the Z”-Score model that
we will subject to empirical analysis. In the fifth section, we discuss the
empirical data and statistical methods, while the sixth section presents
empirical findings. Finally, the seventh section summarizes the study.
2. Classic Z-Score Models
2.1. Z-Score Model for Public Firms
Altman’s (1968) initial sample was composed of 66 corporations, with
33 firms in each of two groups. The bankrupt group (Group 1) con-
134 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
© 2016 John Wiley & Sons Ltd
sisted of manufacturers that filed bankruptcy petitions under Chapter
X of the National Bankruptcy Act during the 1946–1965 period. The mean asset size of these firms was 6.4 million USD, ranging between
0.7 and 25.9 million USD. Altman recognized that this group was not
homogenous with respect to size and industry, although all firms were
relatively small and from manufacturing industries. He attempted to
carefully select non-bankrupt firms (Group 2). Group 2 consisted of a
paired sample of manufacturing firms chosen on a stratified random
basis. These firms were stratified by industry and size, with the asset
size range restricted to 1–25 million USD. Altman eliminated small firms (less than 1 million U.S.A. dollars in total assets) because of a
lack of data and very large firms because of the rarity of bankruptcies
among these firms in that period. He did not match the asset size of
the two groups exactly, and therefore, the firms in Group 2 were
slightly larger than those in Group 1. The data collected for the firms
in both groups were from the same years. For Group 1, the data were
derived from financial statements one reporting period prior to bank-
ruptcy.
Using financial statements, Altman compiled a list of 22 potentially
important financial ratios for evaluation. He classified these variables
into five standard ratio categories: liquidity, profitability, leverage, sol-
vency, and activity. These ratios were chosen based on their popularity
in the literature and their potential relevance to the study. The final
discriminant function estimated by Altman (1968) is as follows:
Z ¼ 0:012�X1 þ 0:014�X2 þ 0:033�X3 þ 0:006�X4 þ 0:999�X5 ð1Þ
where X1 = Working Capital/Total Assets; X2 = Retained Earnings/ Total Assets; X3 = Earnings before Interest and Taxes/Total Assets; X4 = Market Value of Equity/Book Value of Total Liabilities; X5 = Sales/Total Assets; Z = Overall Index.
2.2. Z’-Score and Z”-Score Models for Private Firms
The original Z-Score model was based on the market value of the firm
and was thus applicable only to publicly traded companies. Altman
(1983) emphasized that the Z-Score model is intended for publicly
traded firms and that ad hoc adjustments are not scientifically valid.
Altman (1983) advocated a complete re-estimation of the model, sub-
Financial Distress Prediction 135
© 2016 John Wiley & Sons Ltd
stituting the book value of equity for the market value in X4. Using
the same data, Altman extracted the following revised Z’-Score model:
Z0 ¼ 0:717�X1 þ 0:847�X2 þ 3:107�X3 þ 0:420�X4 þ 0:998�X5 ð2Þ
where X4 = Book value of equity/Book value of total liabilities, with the other variables the same as those in the original (1968) Z-Score
model.
Due to the lack of a private firm database, Altman did not test the
Z’-Score model on a secondary sample. However, he analyzed the
accuracy of a four-variable Z”-Score model that excluded the Sales/
Total assets ratio, X5, from the revised model because of a potential
industry effect that is more likely to take place when this kind of
industry-sensitive variable (asset turnover) is included in the model.
Altman then estimated the following four-variable Z”-Score model
(Altman, 1983):
Z00 ¼ 3:25 þ 6:56�X1 þ 3:26�X2 þ 6:72�X3 þ 1:05�X4 ð3Þ
The EBIT/Total assets ratio, X3, contributed most to the discrimina-
tion power in this version of the model. The classification results for
the Z”-Score model were identical to the revised five-variable Z’-Score
model. In the current study, our empirical analysis focuses on the per-
formance of the Z”-Score model version in predicting bankruptcy,
where it has its widest scope, as it is intended for both privately held
and publicly listed firms and for both manufacturing and non-manu-
facturing firms.
3. Survey of Literature Related to the Altman Z-Score Model
We focus on papers published after 2000 in prominent international
journals and books. 5 Of the many articles and books identified, we
selected 31 articles in which the Z-Score was either used as a failure
prediction proxy or assessed mostly in terms of predictive ability. Of
the 31 studies, Altman’s Z-Score model was used in 16 cases as the
measure of distress or of financial strength. 6 In 13 studies, Altman’s
original model was modified and (or) verified, including re-estimation,
and in two cases, it was used solely for the robustness check. As Pin-
dado et al. (2008) noted, the Z-Score was also used for other purposes,
136 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
© 2016 John Wiley & Sons Ltd
such as the evaluation of the costs and benefits of covenants in bonds,
the choice of debt type (bank versus non-bank, private or public), and
the relationship between investment and internal funds. We focused on
this part of the literature that verified and (or) modified Altman’s
original model. The broad use of the Z-Score model for measuring
financial distress and performing robustness checks indicates its accept-
ability as a reasonable, simple, and consistent measure of distressed
firms.
The most common modification to the Z-Score model was the use
of other estimation techniques or country-specific data. The use of
Altman’s ratios in combination with techniques other than MDA
improved its prediction capability. The application of new data on
both United States and non-US firms also improved model perfor-
mance. Kwak et al. (2005) used Multiple Criteria Linear Programming
(MCLP) to model 5 Altman and 9 Ohlson variables with data on
bankrupt US firms from 1992 to 1998 and nearly six times more
matched U.S.A. control firms. The MCLP approach performed better
than Altman’s original model and gave results similar to or better than
those of Ohlson’s original model. The original models were not recal-
culated, and the authors referred solely to their original prediction
rates.
Merkevicius et al. (2006), using data on United States and Lithua-
nian firms, developed a hybrid artificial discriminant model combining
MDA and an unsupervised learning artificial neural network. This
hybrid SOM-Altman model reached a high prediction rate of 92.35 per
cent. Xu and Zhang (2009) applied Altman’s Z-score, Ohlson’s O-
score, and Merton’s distance-to-default (D-score) models to Japanese
firms to check whether these models are useful for bankruptcy predic-
tion in Japan. They also “merged” these models into a new C-score
model. They then introduced variables unique to Japan to check
whether corporate structure variables have any impact on the probabil-
ity of bankruptcy; they called this the X-score model. These two mod-
els were useful for Japanese firms in predicting bankruptcy, but the
market-based model was the most successful. In summary, the C-score
and X-score (with country-specific variables) models improved bank-
ruptcy prediction.
Tinoco and Wilson (2013) used the original Z-Score as one of the
benchmarks to assess the performance of their model developed for
U.K. listed companies with combined accounting, market, and macroe-
conomic data. Altman’s Z-Score presented very good classification
Financial Distress Prediction 137
© 2016 John Wiley & Sons Ltd
accuracy in the case of financially distressed firms (81 versus 87 per
cent for the new model); however, it was less correct for non-distressed
firm’s prediction. Another modification was introduced by Lyandres
and Zhdanov (2013), who posed the question of whether the inclusion
of variables related to investment opportunities improved the predic-
tive power of three models (Altman’s Z-score model and Zmijewski’s
and Shumway’s models). They used three proxies for investment
opportunities (market-to-book, value-to-book, and R&D-to-assets).
The measures of investment opportunities were linked to the likelihood
of default. The inclusion of either of these measures improved the out-
of-sample forecasting ability of all three.
The verification of Altman’s model concentrated on its efficacy or
on how it compares with other accounting-based, market-based, or
hazard models. However, although the original Z-Score model was not
solely based on accounting data because the market value of equity
was utilized, we classify it here as accounting-based. Grice and Ingram
(2001) used a novel dataset of US firms and posed three questions
about the efficacy of Altman’s model, concluding that the prediction
accuracy of Altman’s model had declined over time and that the coeffi-
cients of the model had significantly changed, which means that the
relation between the financial ratios and the signs of financial distress
had changed over time. The model was sensitive to industry classifica-
tion (more efficient for manufacturing firms than for non-manufactur-
ing firms) but was not sensitive to the type of financial distress. Similar
conclusions were drawn by Grice and Dugan (2003) regarding Ohl-
son’s (1980) and Zmijewski’s (1984) models.
Hillegeist et al. (2004) compared Altman’s Z-score and Ohlson’s O-
score (with original and updated coefficients) with a model based on
Black-Scholes-Merton (BSM) option pricing (a so-called BSM-Prob
model). Hillegeist et al. used relative information content tests to com-
pare the out-of-sample performance of these various models and deter-
mined that BSM-Prob outperformed the alternative accounting-based
models. The conclusions were robust to various modifications of
accounting-based models, such as updated coefficients, industry effects,
and the separation of variables. Chava and Jarrow (2004) employed an
extended bankruptcy database of U.S.A. listed firms to test the superi-
ority of Shumway’s model (2001) over Altman’s (1968) and Zmijew-
ski’s (1984) models. The authors re-estimated the models over the
1962–1990 period and forecasted bankruptcies over the 1991–1999 per- iod. In the case of Shumway’s model, 74.4 per cent (in the first decile)
138 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
© 2016 John Wiley & Sons Ltd
of the bankruptcies were correctly identified; with Altman’s model,
63.2 per cent; and with Zmijewski’s model, 43.2 per cent. Shumway’s
market-based model also outperformed accounting-based models in
terms of the ROC curve (0.91).
Reisz and Perlich (2007) developed a model incorporating barrier
options for bankruptcy prediction and compared its discriminatory
power with other market-based models and Altman’s Z-Score and
Z”-Score. The dataset covered nearly 6000 industrial firms over the
1988–2002 period. The authors documented the superiority of Alt- man’s Z-Score and Z”-Score models for short-term (up to 1 year)
bankruptcy prediction. For medium- and long-term bankruptcy predic-
tion, their barrier option model outperformed the other models.
Pindado et al. (2008) developed an ex ante model for the estimation
of financial distress likelihood (FDL) using a panel data methodology
and presented a financially (not legally) based definition of distress.
Their sample covered 1,583 U.S.A. companies and 2,250 companies
from other G-7 countries for the 1990–2002 period. They used a re- estimated Z-Score as a benchmark. The FDL model outperformed the
Z-Score model in terms of stability and classification power for differ-
ent countries and periods. In the case of the re-estimated Z-Score
model, only profitability and retained earnings maintained their signifi-
cance for different years and countries.
Wu et al. (2010) evaluated the performance of five models (Altman,
1968; Ohlson, 1980 Zmijewski, 1984; Shumway, 2001; Hillegeist et al.,
2004) using an up-to-date dataset for U.S.A. listed firms. Based on
these models, the authors built their own integrated model, that is, a
multi-period logit model with an expanded set of variables. The inte-
grated model, which combined accounting and market data, as well as
firms’ characteristics, outperformed the other models. Altman’s Z-
score performed poorly compared with the other four models. Shum-
way’s model performed best, Hillegeist et al.’s model performed ade-
quately, and Ohlson’s and Zmijewski’s models performed adequately,
although their performance deteriorated over time.
Jackson and Wood (2013) tested 13 different models of bankruptcy
prediction and assessed their efficacy using the ROC curve. They
selected three single-variable models, three accounting-based models
(including Altman’s Z-Score) in two versions (with updated coefficients
and a neural network approach), and four contingent claims models,
and the latter group outperformed the other models. The four best
models were contingent claims models based on European call and
Financial Distress Prediction 139
© 2016 John Wiley & Sons Ltd
barrier options. Although the predictive performance improved with
the application of the neural network to accounting-based models, it
was still lower than with the market-based models.
Acosta-Gonz�alez and Fern�andez-Rodr�ıguez (2014) used genetic algorithms with the Schwarz information criterion (GASIC) for vari-
able selection combined with the logit model for bankruptcy predic-
tion. Altman’s Z-Score model was used as one of two benchmarks
for the authors’ model evaluation. For one-step-ahead forecasting,
Altman’s model was better at predicting failed firms, but the type II
error was high. For two- and three-steps-ahead forecasts, the perfor-
mance of the models was similar for failed firms, but for non-failed
firms, and the prediction accuracy of Altman’s model was worse. For
four-steps-ahead forecasts, the GASIC model outperformed the other
models for failed firms, but it performed comparably for non-failed
firms.
In general, for the 31 articles we reviewed, Altman’s Z-Score model
underperformed compared with market-based models, but evidence
indicated that it performed well for short-term distress prediction. The
question of whether market-based models perform better than account-
ing-based models has been raised many times (e.g., discussion in Das
et al., 2009; Bauer and Agarwal, 2014). Our purpose is not to con-
tribute to this strand of research but rather to focus on the account-
ing-based approach. In this study, we primarily analyze privately held
firms; by definition, there are no market data for these firms. In this
case, an accounting-based approach is the only solution applicable by
banks as lenders, or by investors holding debt securities of firms not
listed on the stock exchange. Thus far, most studies have concentrated
on the U.S.A. market; only a few of them have used data from other
countries, such as Japan, the U.K., Lithuania, or the G-7 countries.
Our analysis is a significant extension of the previous research.
4. Methodology and Research Hypotheses
The literature survey shows that the Z-Score model (publicly traded
firms), the Z’-Score model (private manufacturing firms), and the Z”-
Score model (private and publicly traded manufacturing and non-man-
ufacturing firms) have been adapted for different purposes. In this
study, we are interested first in assessing the performance of the origi-
nal Z”-Score model in classifying bankrupt and non-bankrupt firms in
an international context, with special focus on the European market. © 2016 John Wiley & Sons Ltd
140 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
However, we also validate the results in a set of non-European coun-
tries to generalize the results for circumstances outside Europe. Sec-
ond, we re-estimate the model using extensive international data and
then use the re-estimated Z”-Score model as a benchmark for assessing
the effects of different factors on the model’s performance in terms of
classification accuracy. We assess the effects of the following five fac-
tors on this performance: year of bankruptcy, size of firms, age of
firms, industry, and country of origin. We test a set of hypotheses
based on the effects of the model on performance on two different
levels. First, we test a set of hypotheses on a pooled set of all firms
and, second, on data from each country individually. Compared with
previous research, this study’s contribution is its focus on an inter-
national context, not simply model application or re-estimation of a
given country’s data. Because we focus on the performance of the
Z”-Score model when using a large body of international data, the
research hypotheses are of a technical nature and are given as
follows.
4.1. H1: Obsolescence of the Coefficients
The Z”-Score model was originally estimated using the same sample of
firms used to develop the Z-Score model. The bankruptcies in the esti-
mation data occurred during the 1946–1965 period. Thus, the oldest observations are from nearly seventy years ago, during the post-war
period. Altman (1983) recommended utilizing data as near to the pre-
sent as possible when developing a bankruptcy prediction model. It is
obvious that firms’ financial behavior and their business environment
have significantly changed since then. The importance of the financial
ratios, as reflected by the coefficients of the model, may differ from
their original importance. Therefore, we suggest, as the first hypothesis
(H1), that the re-estimation of the coefficients of the four original vari-
ables of the Z”-Score model will improve the classification perfor-
mance of the model in an international context. This hypothesis is
supported by the previous research (e.g., Grice and Ingram, 2001) and
by practice. H1 aims to support this evidence on the international level
and is expressed as follows:
H1: Re-estimating the coefficients of the Z”-Score model improves its clas-
sification accuracy.
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 141
4.2. H2: Method of Estimation
The original Z”-Score model has been estimated using MDA. How-
ever, MDA is based on the ordinary least squares (OLS) method and
thus requires assumptions of multinormality, homoscedasticity, and
linearity, which are not often met in empirical financial ratio analysis.
We re-estimate the Z”-Score model using logistic regression analysis
(LRA) to assess the effect of the estimation method. LRA does not
require most of the restricting assumptions of MDA. In LRA, the
multivariate normality of the independent variables is not required,
nor are homoscedasticity and linearity. For the sake of OLS, MDA
can be more useful than LRA for small samples, such as the original
sample of 66 firms used in the estimation of the Z”-Score model. How-
ever, in a large sample, LRA may potentially perform better. In this
study, we use large samples, which is advantageous for LRA. Our sec-
ond hypothesis (H2) is that the classification performance of the re-
estimated Z”-Score model will improve when it is estimated using
LRA instead of MDA. The model re-estimated for the original vari-
ables using LRA and all pooled data arecalled the Z”-Score LR
model. The performance of this re-estimated model is used as the
benchmark for further analyses. Thus, the second hypothesis is as fol-
lows:
H2: The prediction accuracy of the logistic regression version of the
Z”-Score model is higher than that of the multiple discriminant analysis
version.
4.3. H3: Bankruptcy Year
The model based on the relationship between bankruptcy and financial
ratios is likely to be affected by the macroeconomic environment.
These effects may significantly decrease the classification accuracy of
the model. If the model is estimated using data from 1 year and will
be applied to data from another year, the validity of the model can be
questioned. In terms of economic growth, credit policy, and interest
rates, business cycles can affect the boundary between bankrupt and
non-bankrupt firms. The original Z”-Score model is estimated using
data from the 1946–1965 period, which includes several business cycles. Therefore, the model is not focused on any specific stage of a cycle
© 2016 John Wiley & Sons Ltd
142 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
and does not explicitly take into account the bankruptcy year. Altman
(1983) suggested gathering data from firms for the most recent few
years when developing a prediction model. In this study, the bench-
mark Z”-Score LR model is estimated for a shorter period than in the
original estimation, although it covers several recent years for different
stages of the business cycle in different countries. The third hypothesis
(H3) is that the classification accuracy of the benchmark model can be
increased by explicitly taking into account the year of bankruptcy in
the model estimation. The third hypothesis is as follows:
H3: The model’s prediction accuracy is higher when the effect of the year
of bankruptcy is included.
4.4. H4: Size of the Firm
The boundary between bankrupt and non-bankrupt firms is different for
small and large firms, which decreases the performance of the model
estimation when data from one size category are applied to another size
category. For the bankrupt and non-bankrupt firms in the original data
for Z”-Score model estimation, asset sizes ranged between approxi-
mately 1 and 25 million U.S.A. dollars. The data did not include very
small or very large firms. Altman (1983) regarded the suitability of the
original Z-Score model (and, likewise, the Z”-Score model) for all firms
as debatable because of this omission. In the current study, the bench-
mark Z”-Score LR model is estimated for data from many size cate-
gories, from very small to very large firms. The fourth hypothesis (H4)
assumes that the classification performance of the uniform benchmark
LR model based on the original four financial variables of the Z”-Score
model is improved when the size category of the firm is explicitly taken
into account. Thus, we present the fourth hypothesis:
H4: The model’s prediction accuracy is higher when the effect of size is
included.
4.5. H5: Age of the Firm
International insolvency statistics generally show that bankruptcy risk
is a function of the age of the firm. Very young firms typically show
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 143
very high risk. The original Z”-Score model does not explicitly take
age into account. However, Altman (1983) noted that the age of a firm
is implicitly considered in the Retained Earnings/Total Assets ratio
(X2), which was regarded as a new ratio in the bankruptcy prediction
context. A relatively young firm will probably show a low ratio
because it has not had time to build up cumulative profits. Thus, a
young firm is, to some degree, discriminated against in the model, and
its likelihood of being classified as bankrupt is relatively higher than
that of an older firm. The incidence of failure is much higher in the
early years of a firm. Although the age of the firm is implicitly taken
into account in X2, we expect that an explicit consideration of age will
improve the classification accuracy by controlling for the age factor.
The fifth hypothesis (H5) proposes that the performance of the uni-
form benchmark model based on the original four financial variables
of the Z”-Score model increases when the age of the firm is explicitly
taken into account. The fifth hypothesis is as follows:
H5: The model’s prediction accuracy is higher when the effect of firm age
is included.
4.6. H6: Industry of the Firm
The original Z’-Score model was estimated only for manufacturing
firms. Altman (1983) stated that it would be ideal to develop a bank-
ruptcy prediction model utilizing a homogenous group of bankrupt
firms. If we are interested in a particular industry grouping, we should
gather data from bankrupt and non-bankrupt firms in that grouping.
Previous studies show that financial distress analysis is influenced by
the industry effect (Smith and Liou, 2007). Firms in different industries
tend to report different levels of the same financial ratios, which may
have an effect on the boundary between bankrupt and non-bankrupt
firms. This industry effect may be present in the Z’-Score model, espe-
cially due to the Sales/Total Assets ratio (X5), which showed the least
significance on a univariate basis while making a very significant con-
tribution to the discriminant power of the multivariate model. Altman
(1983) recognized the potential industry effect due to the wide varia-
tion among industries in asset turnover and specified the Z”-Score
model without X5. However, the Z”-Score model was also estimated
using the original sample of manufacturing firms. In our analysis, the
© 2016 John Wiley & Sons Ltd
144 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
uniform benchmark model based on the original four financial vari-
ables of the Z”-Score model is estimated for a statistical sample repre-
senting different industries. The sixth hypothesis (H6) assumes that an
explicit consideration of industry will improve the classification accu-
racy of this benchmark model. H6 can be expressed in the following
form:
H6: The model’s prediction accuracy is higher when the effect of industry
is included.
4.7. H7: Country of Origin
The original Z”-Score model has been estimated only for U.S.A. firms.
It can be expected that the international applicability of the model to
other countries is affected by country-specific differences. The eco-
nomic environment, legislation, culture, financial markets, and
accounting practices in a country may affect the financial behavior of
firms and the boundary between bankrupt and non-bankrupt firms.
These factors may weaken the classification performance of the model
for countries other than that for which the model was originally esti-
mated (Ooghe and Balcaen, 2007). The seventh hypothesis (H7)
assumes that explicitly taking the country of origin of a firm into
account will improve the classification accuracy of the benchmark
model. In our empirical study, the country effect is assessed by includ-
ing a variable for country risk. The seventh hypothesis is as follows:
H7: The model’s prediction accuracy is higher when the effect of country
risk is included.
5. Empirical Data and Statistical Methods
5.1. Sample of Firms
The principal data for this study were extracted from the ORBIS data-
bases of Bureau Van Dijk (BvD). ORBIS Europe is a commercial
database that, at the moment of sampling, contained administrative
information on more than 50 million European firms. However,
income statement and balance sheet information was available for
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 145
approximately 8 million companies. More than 99 per cent of the com-
panies covered in this database are private companies from various
industries, justifying the use of the Z”-Score model instead of the orig-
inal Z-Score model. The Z”-Score model was originally made robust
across all industrial groupings and for both private and public entities
(Altman, 1983, 2014; Altman and Hotchkiss, 2006). Because we do not
want to limit the scope of the Z”-Score model in this study, we retain
both private and public firms from all industrial groupings.
The ORBIS formats have been derived from the world’s most com-
monly used formats for the presentation of business accounts (Ribeiro
et al., 2010). International comparability may be a problem when
administrative firm-level data are pooled across countries. Although
the definition of variables is usually less harmonized for administrative
data, this is less of a problem in the ORBIS database because of the
common international format of balance sheets. For example, although
some discrepancies in profit/loss statements may arise because of dif-
ferences in fiscal systems across countries, balance sheet variables lar-
gely adhere to international standards.
A number of factors influence the international applicability of
bankruptcy prediction models: accounting legislation and practice,
creditor rights and investor protection, judicial efficiency, corporate
governance, bankruptcy protection and insolvency management, and
firm risk-taking. These factors strongly differ between European and
non-European countries. Therefore, we aim to test the performance of
the Z”-Score model outside Europe. First, it is particularly important
to include the US because it is the country of origin for the Z”-Score
model and because it has the largest market capitalization in the
world. Second, a central motivation in developing the modified Z”-
Score model was to make it applicable to emerging market companies.
We include firms from China and Colombia, which represent two very
culturally and institutionally dissimilar emerging market countries. For
other non-European countries, sufficient bankruptcy data (more than
60 bankrupt firms, i.e., the limit we set for European countries) from
ORBIS World were not available. Thus, the results are also estimated
and tested for three non-European countries (the United States, China,
and Colombia) to gain a more global view of the Z”-Score model’s
performance. The samples of firms from these countries were extracted
from ORBIS World, which contains middle-sized (total assets over 1.5
million EUR) and larger firms from around the world.
© 2016 John Wiley & Sons Ltd
146 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
Several requirements are set for the statistical sampling of the
empirical data. First, we require that the firm to be selected be an
industrial (non-financial) company. Second, its owners must have lim-
ited liability (so partnerships and sole proprietors are left out of the
study). Third, we set a minimum requirement for the size of the firm.
Because the financial ratios of very small firms are generally too unsta-
ble for a failure prediction model, these firms are excluded (see Balcaen
and Ooghe, 2006). We require that a firm’s total assets must have
exceeded 100 thousand EUR at least once in the available time series.
Fourth, we include in our estimation sample firms from all European
countries and three pre-selected non-European countries where the
number of failed firms is greater than 60. If the number of failed firms
for any European country is less than 60, the firms from this country
are included only in the test sample. For qualifying European coun-
tries, firms are randomly classified in the estimation and test samples
so that the number of firms is approximately equal in both samples.
Thus, our estimation data include firms from 28 European and three
non-European countries. Fifth, all failed firms that fulfill the above
requirements are included in our samples. However, if the number of
non-failed firms in a country is very high, a sample is randomly
selected from that country. Finally, the time span of fiscal years poten-
tially available for this study ranges from 2002 to 2010. Because the
most recent financial statements for failed firms in the database are
from a financial period within 2007 and 2010, earlier years are also
excluded for non-failed firms for comparability. All qualifying observa-
tions of non-failed firms from 2007 to 2010 are included in the data-
sets. We restrict the analyses of failed firms to the most recent
financial statements available before failure. The four independent
variables of the Z”-Score model were winsorized at 1 and 99 per cent
to minimize outliers.
Table 1 shows the resulting number of non-failed and failed firms in
the estimation data and test data by country. The estimation sample
includes data from 2,602,563 non-failed and 38,215 failed firms from
28 European and three non-European countries. The test sample is
slightly larger because it includes data from 31 European and three
non-European countries. For the country of origin of the Z”-Score
model, the United States, the estimation sample includes only 56 bank-
rupt firms. The available U.S.A. data consist only of listed (and
delisted) firms. From China, there are three sub-samples. Only 32 pub-
lic firms with special treatment (ST) 7 status are included in the estima-
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 147
tion sample. 8 The Chinese datasets of predominantly private firms
(CN) and of public firms with delisted (DL) failure status are analyzed
separately only for the test data. 9 ST firms are listed firms suffering
from serious financial difficulties. Excluding the special U.S.A. data
and the two non-private Chinese datasets, 99.4 per cent of the observa-
tions in the data are private firms.
Table 1. Number of Observations by Country
Country
Estimation data Test data
Non-failed Failed Non-failed Failed
Austria (AT) 7,430 55 7,526 44 Belgium (BE) 179,979 2,994 179,818 2,944 Bosnia (BA) 29,391 35 29,139 32 Bulgaria (BG) 50,041 48 42,351 44 China (CN) 39,315 198 China, delisted dataset, DL 16,291 29 China, ST data 846 16 1,020 16 Colombia (CO) 8,366 139 6,982 125 Greece (GR) 51,763 28 Croatia (HR) 59,541 249 58,478 275 Czech Republic (CZ) 92,835 556 92,562 564 Denmark (DK) 167,934 1,334 168,538 1,398 Estonia (EE) 34,313 234 34,196 242 Finland (FI) 90,878 481 91,227 459 France (FR) 160,749 6,124 161,653 6,318 Germany (DE) 98,814 910 99,496 921 Hungary (HU) 19,421 303 20,155 313 Iceland (IS) 17,399 248 17,624 243 Ireland (IE) 6,665 121 6,406 139 Italy (IT) 167,113 8,101 166,258 8,124 Latvia (LV) 8,064 433 8,241 477 Lithuania (LT) 10,000 56 Netherlands (NL) 20,885 154 15,854 147 Norway (NO) 172,467 1,294 170,985 1,206 Poland (PL) 87,200 291 86,233 264 Portugal (PT) 180,114 3,390 178,646 3,422 Romania (RO) 161,992 97 164,259 93 Russian Federation (RU) 116,903 2,534 115,711 2,481 Serbia (RS) 100,100 68 Slovakia (SK) 7,856 120 7,788 124 Slovenia (SI) 14,419 59 14,081 41 Spain (ES) 156,746 3,036 158,122 2,991 Sweden (SE) 169,810 2,256 169,999 2,314 Ukraine (UA) 133,342 1,787 133,980 1,765 United Kingdom (GB) 171,493 760 170,930 716 U.K. (GB), liquidation dataset 342,423 4,990 United States (US) 9,557 56 9,929 53 Total 2,602,563 38,215 3,148,079 43,664
© 2016 John Wiley & Sons Ltd
148 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
5.2. Status of Failed Firms
ORBIS has five classes for potentially active firms (active, default of
payment, receivership, dormant, and branch) and seven classes for
inactive firms that no longer carry out business activities (bankruptcy,
dissolved, dissolved-merger, dissolved-demerger, in liquidation, branch,
and no precision). Among these classes, only active is selected to repre-
sent non-distressed firms. In selecting the failed firms, we try to avoid
ambiguity by considering (with exceptions described below) a firm
failed if its status in ORBIS is stated as bankruptcy. However, because
of the small number of bankrupt firms in some countries, we also con-
sider receivership (active) firms as failed even if they are active. These
firms generally suffer from serious financial distress. However, firms in
liquidation are generally not included in the sample of failed firms.
Firms in liquidation may, depending on the country, contain firms that
have ceased activities due to reasons other than failure (mergers, dis-
continuing the operations of a daughter company or of a foreign
branch, etc.). Therefore, for most countries, we select only firms that
are coded as being bankrupt or under receivership. However, there are
a number of special cases where failed firms are coded under a differ-
ent status heading. These countries or samples are the following:
Country Status categories
Bulgaria In liquidation, Bankruptcy Denmark Inactive (no precision) Greece Active (receivership), In liquidation, Bankruptcy Ireland In liquidation, Active (receivership) Norway In liquidation Slovenia In liquidation Spain Active (receivership), In liquidation, Bankruptcy Ukraine In liquidation, Bankruptcy U.K., liquidation set In liquidation China, ST Active (special treatment) China, delisted, DL Active (delisted)
If no such category of failed firms could be identified, that country
was excluded from the study (for example, Switzerland). If a country
had only a very small number of failed firms, it was dropped from the
study (typically small countries, including Luxembourg, Liechtenstein,
and Montenegro). It should also be noted that the status classes (in-
cluding the bankruptcy category) are not completely homogenous
within European countries due to different legislations, although there
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 149
are obvious similarities in insolvency acts (Philippe et al., 2002). China
is a special case because it includes samples with three different criteria
of failure (bankruptcy, special treatment, and delisted). Additionally,
for the United Kingdom, there are two different samples (liquidation
and receivership).
5.3. Statistical Methods
In this study, seven research hypotheses are drawn for statistical
testing. The statistical analysis begins with calculating the original
Z”-Score for the firms in the data, as in equation (3). The classification
performance of the original model is assessed by the AUC (Area Under
Curve) measure extracted from the ROC curve. AUC has a close con-
nection with the Accuracy Ratio (AR) because AR = 2 � AUC – 1. AR equals 0 for a random model, 1 for a perfect model, and 0.5 for a
model with an average classification performance. SAS software (SAS
Institute Inc., Cary, NC, USA) is used for all statistical analyses.
The first hypothesis (H1) assumes that the coefficients of the original
model are obsolete. H1 is tested by re-estimating the coefficients of the
Z”-Score model using the original statistical method (multiple discrimi-
nant analysis, or MDA). The problem is that the estimation sample
includes different numbers of failed and non-failed firms from 31 coun-
tries. In the original Z-Score” sample (1983), equal numbers of bank-
rupt and non-bankrupt firms were selected from the US Following the
characteristics of these data. Therefore, we weight the failed and non-
failed firms equally. In so doing, the non-proportional sampling from
different countries will not affect the re-estimated model. The number
of firms from different countries, however, varies significantly, leading
to greater weights for larger countries. To avoid this problem, the
observations are also weighted so that each country has an equal
weight in the analysis. Then, the coefficients of the Z”-Score model are
re-estimated using these weighted data, and the resulting AUC is com-
pared with that based on the original model.
The second hypothesis (H2) tests whether the classification perfor-
mance of the re-estimated Z”-Score model improves when it is re-esti-
mated using logistic regression analysis (LRA), which is based on less-
restrictive statistical assumptions than MDA. In this estimation, the
dependent variable Y = 0 for non-failed firms and Y = 1 for failed firms. LRA does not require that independent variables be multivariate
normal or that groups have equal covariance matrices, which are basic © 2016 John Wiley & Sons Ltd
150 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
assumptions in MDA (Hosmer and Lemeshow, 1989). LRA creates a
(logit) score L for every firm. It is assumed that the independent vari-
ables are linearly related to L. This score or logit is used to determine
the conditional probability of failure as follows:
pðY ¼ 1jXÞ ¼ 1 1 þ e�L ¼
1
1 þ e�ðb0þb1X1þ…þb4X4Þ ð4Þ
where bi (i = 0, . . . 4) are the coefficients and Xi (i = 1,. . ., 4) are the four independent variables of the original Z”-Model.
10 The effect of
this method on classification performance is assessed by testing the sta-
tistical significance of the difference between AUCs for this LR model
and for the re-estimated MDA model. The resulting model is called
the Z”-Score LR model, and it is used as a benchmark for further sta-
tistical AUC comparisons because LR is applied as the principal
method in testing the remaining research hypotheses.
The third hypothesis (H3) is associated with the performance effect
of taking account of the bankruptcy year in the estimation. This
hypothesis is tested by estimating an LR model based on the following
logit:
L ¼ b0 þ X4
i¼1 biXi þ
X3
j¼1 cjDj ð5Þ
where b0 is a constant, Xi (i = 1,. . ., 4) are the four independent vari- ables of the original Z”-Model, bi (i = 1,. . ., 4) are their coefficients, cj (j = 1,. . ., 3) are coefficients of the dummy variables, and D1 = 1 when year = 2008, 0 otherwise; D2 = 1 when year = 2009, 0 otherwise; D3 = 1 when year = 2010, 0 otherwise.
The dummy variables do not directly refer to the bankruptcy year,
which is not given in the database, but rather to the last available year
before bankruptcy. For failed firms, there is an approximately 1–2-year lead time to failure from this year. In this model, the year 2007 is the
base category. If the AUC of this extended LR model statistically sig-
nificantly exceeds the AUC of the Z”-Score LR model (benchmark),
the evidence supports hypothesis H3.
Research hypotheses H4–H7 are tested using the same approach as the third hypothesis above. However, for each hypothesis, appropriate
variables are used instead of the year dummies. Hypothesis H4
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 151
addresses the performance effect of taking size into account, and it is
tested by adding two additional variables measuring firm size to the
LR model. In this LR model, size is measured by the natural loga-
rithm of total assets and its squared form. In this way, the effect of
logarithmic size can be reflected in a function following the second-
order parabola. Hypothesis H5 tests whether the classification perfor-
mance improves when the age of the firm is explicitly taken into
account. When testing this hypothesis, the 6–14 years category is used as the base category, and two dummy variables are incorporated in the
LR model (D1: less than 6 years, D2: 15 years or more). Hypothesis
H6 looks at whether classification performance is affected by the expli-
cit consideration of industry effects. The hypothesis is tested here using
dummy variables for seven industries (D1: restaurants and hotels; D2:
construction; D3: wholesale and retailing; D4: agriculture; D5: manufac-
turing; D6: energy and water production; D7: information technology),
with all other industries acting as the base category.
Hypothesis H7 tests whether the explicit consideration of the coun-
try of origin will improve classification performance. This hypothesis is
tested by using country risk measures instead of dummy variables for
countries. The country risk of each country is measured by Standard
& Poor’s Country Risk Rating per 6 months after the annual closing
of accounts. The rating is numerically recoded such that the best rat-
ing, AAA, equals 1, the second-best rating, AA+, equals 2, and so on. Finally, the lowest rating, D, equals 22. Thus, H7 is tested by estimat-
ing an LR model based on the four financial ratios of the original Z”-
Score model and a 22-step variable referring to country risk. The five
LR models with the original four financial ratios and the additional
variables specified in the hypotheses are estimated for all data. In addi-
tion, an LR model including all additional variables is estimated for
all data to assess the simultaneous effect of all variables. Finally, six of
the seven hypotheses are tested for the data of each country separately.
Hypothesis H7 is not included in this country-level testing because the
additional variable (country risk) is constant within the country.
6. Empirical Results
6.1. All Data: Coefficients of the Z”-Score Models
Table 2 presents descriptive statistics of the four independent variables
(X1–X4) of the Z”-Score model for all data. The variation in the ratios © 2016 John Wiley & Sons Ltd
152 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
is significant, as shown by the standard deviation and the quartiles.
For X1 (WCTA), X2 (RETA), and X3 (EBITTA), the median and the
mean for non-failed firms are close to each other, indicating a symme-
try of distributions. However, this is not the case for failed firms. For
failed/distressed firms, the median exceeds the mean for these three
ratios, indicating negatively skewed distributions. For X4 (BVETD),
the means significantly exceed the median for both failed and non-
failed firms, indicating a positively skewed distribution. For each of
the four variables, both the mean and the median are higher for non-
failed firms, which is consistent with our expectations. The difference
between the means of non-failed and failed firms is larger in the origi-
nal U.S.A. data than in our all data for Retained Earnings/Total
Assets (RETA) and EBIT/Total Assets (EBITTA) but is approxi-
mately the same size for Working Capital/Total Assets (WCTA) and
Book Value of Equity/Total Liabilities (BVETD) (Altman, 1983).
These characteristics of the data may indicate lower classification accu-
racy than in the original sample.
Table 3 presents the coefficients of the different models estimated
for all data. All LRA estimates (Model 2 to Model 9) are statistically
significant at 0.0001 due to their contributions and to the large sample
size. The first column presents the coefficients of the original Z”-Score
model. The “Model 1” column shows the coefficients when they are
re-estimated by the same statistical method, specifically MDA. The
coefficients here are negative because the models are estimated using
Table 2. Descriptive Statistics (All Data)
WCTA RETA EBITTA BVETD
Statistic Non- failed Failed
Non- failed Failed
Non- failed Failed
Non- failed Failed
Median 0.152 �0.059 0.189 �0.024 0.041 �0.020 0.451 0.025 Mean 0.147 �0.213 0.188 �0.317 0.055 �0.108 3.594 0.703 Standard deviation
0.442 0.604 0.509 0.767 0.227 0.296 11.499 5.712
Upper quartile
0.420 0.142 0.476 0.087 0.131 0.042 1.548 0.215
Lower quartile
�0.040 �0.440 0.011 �0.450 �0.008 �0.208 0.100 �0.240
Maximum 0.956 0.956 0.958 0.958 0.785 0.785 68.606 68.606 Minimum �1.637 �1.637 �2.453 �2.453 �0.828 �0.828 �0.649 �0.649
WCTA, Working Capital/Total Assets; RETA, Retained Earnings/Total Assets; EBITTA, EBIT/Total Assets; BVETD, Book Value of Equity/Total Liabilities.
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 153
T a b le
3 . T h e C o effi
ci en ts
o f th e D iff er en t M o d el s E st im
a te d fo r a ll D a ta
C o effi
ci en ts
fo r d iff er en t st a ti st ic a l m o d el s
V a ri a b le
Z ” -S co re
M o d el
1 M o d el
2 M o d el
3 M o d el
4 M o d el
5 M o d el
6 M o d el
7 M o d el
8
C o n st a n t
3 .2 5
�0 .0 4 2
0 .0 3 5
0 .2 0 7
�1 3 .4 6 6
0 .0 0 7
0 .0 4 8
0 .0 4 9
�1 3 .3 0 2
W C T A
6 .5 6
�0 .5 6 1
�0 .4 9 5
�0 .4 8 3
�0 .4 4 1
�0 .4 8 7
�0 .5 4 0
�0 .4 9 6
�0 .4 5 9
R E T A
3 .2 6
�0 .7 2 4
�0 .8 6 2
�0 .8 9 1
�1 .1 4 6
�0 .8 4 6
�0 .8 5 9
�0 .8 6 3
�1 .1 6 0
E B IT
T A
6 .7 2
�1 .7 9 1
�1 .7 2 1
�1 .7 9 0
�1 .6 1 9
�1 .7 5 7
�1 .6 9 5
�1 .7 1 7
�1 .6 8 2
B V E T D
1 .0 5
�0 .0 2 1
�0 .0 1 7
�0 .0 1 6
�0 .0 1 2
�0 .0 1 7
�0 .0 1 6
�0 .0 1 7
�0 .0 1 3
Y ea r d u m m ie s
Y ea r 2 0 0 8
�0 .0 5 5
�0 .0 3 4
Y ea r 2 0 0 9
�0 .1 7 9
�0 .1 5 0
Y ea r 2 0 1 0
�0 .6 6 6
�0 .6 3 1
S iz e v a ri a b le s
T o ta l a ss et s (l o g )
1 .8 3 0
1 .8 3 7
T o ta l a ss et s sq u a re d (l o g )
�0 .0 6 1
�0 .0 6 1
A g e d u m m ie s
L es s th a n 6 y ea rs
0 .1 3 5
0 .1 8 6
1 5 y ea rs
o r m o re
�0 .0 5 8
�0 .0 9 9
C o u n tr y ri sk
S P co u n tr y ra ti n g ra n k
�0 .0 0 3
�0 .0 1 4
In d u st ry
d u m m ie s
R es ta u ra n ts
a n d h o te ls
�0 .6 5 3
�0 .6 2 8
C o n st ru ct io n
0 .4 4 5
0 .3 6 5
W h o le sa le
a n d re ta il in g
�0 .1 1 2
�0 .1 5 7
A g ri cu lt u re
�0 .1 8 0
�0 .1 7 6
M a n u fa ct u ri n g
0 .1 3 9
0 .0 9 5
E n er g y a n d w a te r p ro d u ct io n
�0 .4 5 4
�0 .4 7 2
In fo rm
a ti o n te ch n o lo g y
�0 .9 1 3
�0 .9 1 5
S ig n ifi ca n ce : C o effi
ci en ts
a re
a ll st a ti st ic a ll y si g n ifi ca n t a t 0 .0 0 0 1 .
M o d el s:
Z ” -S co re
= O ri g in a l A lt m a n (1 9 8 3 ) Z ” -S co re
M o d el
co effi
ci en ts ; M o d el
1 =
T h e M D A
m o d el ; M o d el
2 =
T h e L R
m o d el ; M o d el
3 =
T h e L R
m o d el
es ti m a te d fo r a ll d a ta
w it h y ea r d u m m ie s;
M o d el
4 = T h e L R
m o d el
es ti m a te d fo r a ll d a ta
w it h si ze
v a ri a b le s;
M o d el
5 =
T h e L R
m o d el
es ti m a te d fo r a ll d a ta
w it h a g e ca te g o ry
d u m m ie s;
M o d el
6 =
T h e L R
m o d el
es ti m a te d fo r a ll d a ta
w it h in d u st ry
d u m m ie s;
M o d el
7 =
T h e
L R
m o d el
es ti m a te d fo r a ll d a ta
w it h co u n tr y ri sk
ra n k in g s; M o d el
8 =
T h e L R
m o d el
es ti m a te d fo r a ll d a ta
w it h a ll v a ri a b le s.
© 2016 John Wiley & Sons Ltd
154 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
Y = 1 for the failed firms. In our all data, EBITTA has a significantly higher relative weight than in the original U.S.A. data, while the
weights of WCTA and BVETD have proportionally decreased. The re-
estimated coefficient of BVETD is very small, indicating a minor effect
on the logit. The “Model 2” column presents the coefficients for the
Z”-Score LR model. These coefficients are directly comparable with
those of the MDA model, as expected for this exceptionally large sam-
ple. For each model, the coefficient of BVETD is very close to zero.
The differences in the coefficients of the original four variables among
the eight LR models (Models 1–8) are small, indicating that the origi- nal four variables and the additional variables are quite independent
of each other.
Table 3 also shows the coefficients of the additional variables in the
LR models. The negative coefficients of the dummy (year) variables of
Model 3 indicate that after 2007 (the base category), this risk of failure
significantly decreased year by year. The base year of 2007 indicates
that a failure emerged during 2008–2009 as a result of the global finan- cial crisis. The crisis played a significant role in the failure of key busi-
nesses and caused a downturn in economic activity, leading to the
2008–2012 recession. The effects were especially felt in Europe. The coefficients of Model 4 for the size variables show that the contribu-
tion of size to the logit (risk measure) reaches its maximum value when
logarithmic total assets are 15 or when total assets are approximately
3.3 million EUR. Model 5 confirms the riskiness of young firms
because the risk of failure is very high for newly founded firms (less
than 6 years old), as shown by the coefficient of the first dummy vari-
able. The coefficients of the industry dummies in Model 6 show that
construction is an exceptionally risky industry, followed by manufac-
turing. For Model 7, the coefficient of the country risk dummy is sta-
tistically significant (because of the large sample) but negative and very
close to zero. Finally, the coefficients of all variables in Model 8 are
directly comparable to those in Models 3–7.
6.2. All Data: Performance of the Z”-Score Models
Table 4 shows the AUCs in the test data for the different “all data”
models by country. Model 1 refers to the original Z”-Score model.
The classification performance of the score at the level of all countries
is fair because AUC = 0.743 refers to AR = 0.486, which is approxi- mately average accuracy (0.5). However, the score gives relatively good
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 155
T a b le
4 . T es t D a ta
A U C s fo r D iff er en t C o u n tr ie s,
B a se d o n A ll D a ta
M o d el
V er si o n s. C o m p a ri so n s a re
w it h th e A U C
o f th e L R
M o d el
E st im
a te d fo r A ll D a ta
( B en ch m a rk ) . T es t d a ta
A U C
fo r d iff er en t m o d el s
C o u n tr y
B en ch m a rk
M o d el
1 M o d el
2 M o d el
3 M o d el
4 M o d el
5 M o d el
6 M o d el
7 M o d el
8
A ll d a ta
0 .7 4 8
0 .7 4 3 —
0 .7 4 5 —
0 .7 5 2 + + + +
0 .7 6 0 + + + +
0 .7 4 8 + + +
0 .7 5 1 + + + +
0 .7 4 9 + + + +
0 .7 7 1 + + + +
C o u n tr ie s w it h es ti m a ti o n d a ta
A u st ri a (A
T )
0 .8 0 0
0 .7 8 8
0 .7 9 7
0 .8 0 5
0 .8 1 8 +
0 .8 1 4 +
0 .7 8 2
0 .8 0 0
0 .8 1 9
B el g iu m
(B E )
0 .7 7 2
0 .7 6 0
0 .7 7 0
0 .7 7 7
0 .7 4 7
0 .7 7 7
0 .7 6 5
0 .7 7 2
0 .7 5 8
B o sn ia
(B A )
0 .8 6 2
0 .8 0 5
0 .8 5 7
0 .7 7 6
0 .8 6 3
0 .8 4 7
0 .8 5 5
0 .8 6 2
0 .7 8 4
B u lg a ri a (B G )
0 .6 8 4
0 .6 3 0
0 .6 8 0
0 .6 5 4
0 .6 9 1
0 .6 8 0
0 .6 5 9
0 .6 8 4
0 .6 3 2
C h in a , S T
co m p a n ie s
0 .9 8 5
0 .9 1 1 —
0 .9 8 3
0 .9 5 8 –
0 .9 7 7
0 .9 7 8
0 .9 8 7
0 .9 8 5
0 .9 6 8
C o lo m b ia
(C O )
0 .7 2 6
0 .7 2 4
0 .7 2 7
0 .7 1 5
0 .7 5 8
0 .7 2 8
0 .7 2 6
0 .7 2 6
0 .7 5 7
C ro a ti a (H
R )
0 .8 4 4
0 .8 1 2
0 .8 3 9
0 .8 0 3
0 .8 3 7
0 .8 3 5
0 .8 3 2
0 .8 4 4
0 .8 0 1
C ze ch
R ep u b li c (C
Z )
0 .8 1 1
0 .8 1 3
0 .8 1 1
0 .8 1 9
0 .8 2 8
0 .8 0 7
0 .8 2 0
0 .8 1 1
0 .8 3 8
D en m a rk
(D K )
0 .8 0 3
0 .7 9 8
0 .8 0 0
0 .7 8 1
0 .8 1 3
0 .8 0 2
0 .8 0 1
0 .8 0 3
0 .7 9 6
E st o n ia
(E E )
0 .8 2 3
0 .8 2 7
0 .8 2 3
0 .8 4 7
0 .8 6 6
0 .8 2 6
0 .8 3 3
0 .8 2 3
0 .8 9 0
F in la n d (F I)
0 .8 6 7
0 .8 6 4
0 .8 6 6
0 .8 3 5
0 .8 6 2
0 .8 7 0
0 .8 7 8
0 .8 6 7
0 .8 5 3
F ra n ce
(F R )
0 .7 3 9
0 .7 2 3
0 .7 3 5
0 .7 4 9
0 .7 7 1
0 .7 4 1
0 .7 6 2
0 .7 3 9
0 .7 9 9
G er m a n y (D
E )
0 .6 7 3
0 .6 5 8
0 .6 6 6
0 .6 9 5
0 .6 5 6
0 .6 8 4
0 .6 7 7
0 .6 7 3
0 .6 8 8
H u n g a ry
(H U )
0 .7 4 2
0 .7 4 6
0 .7 4 0
0 .6 6 0
0 .7 3 8
0 .7 5 5
0 .7 3 5
0 .7 4 2
0 .6 9 6
Ic el a n d (I S )
0 .6 6 4
0 .6 7 4
0 .6 6 6
0 .6 9 4
0 .6 7 8
0 .6 7 3
0 .6 7 2
0 .6 6 4
0 .7 1 6
Ir el a n d (I E )
0 .6 7 9
0 .6 7 2
0 .6 7 6
0 .7 0 8
0 .6 7 7
0 .6 8 1
0 .6 8 8
0 .6 7 9
0 .7 1 2
It a ly
(I T )
0 .8 0 6
0 .7 9 9
0 .8 0 4
0 .8 3 3
0 .8 3 5
0 .8 0 6
0 .7 9 9
0 .8 0 6
0 .8 4 9
L a tv ia
(L V )
0 .6 7 8
0 .6 9 1
0 .6 7 8
0 .7 0 4
0 .6 7 6
0 .6 8 6
0 .6 9 8
0 .6 7 8
0 .7 2 4
N et h er la n d s (N
L )
0 .7 5 2
0 .7 5 4
0 .7 5 0
0 .7 7 5
0 .7 6 9
0 .7 5 4
0 .7 4 6
0 .7 5 2
0 .7 8 7
N o rw
a y (N
O )
0 .7 1 6
0 .6 9 4
0 .7 1 3
0 .6 5 8
0 .6 8 2
0 .7 2 0
0 .7 1 5
0 .7 1 6
0 .6 4 5
P o la n d (P L )
0 .9 0 3
0 .9 0 4
0 .9 0 4
0 .9 0 8
0 .9 0 2
0 .9 0 3
0 .8 9 9
0 .9 0 3
0 .9 0 4
P o rt u g a l (P T )
0 .7 4 1
0 .7 2 4
0 .7 3 6
0 .7 4 9
0 .7 7 3
0 .7 3 8
0 .7 5 5
0 .7 4 1
0 .7 8 5
R o m a n ia
(R O )
0 .7 5 8
0 .7 4 0
0 .7 5 4
0 .7 0 9
0 .7 4 9
0 .7 5 5
0 .7 4 8
0 .7 5 8
0 .7 0 3
R u ss ia n F ed er a ti o n
(R U )
0 .8 1 1
0 .8 0 2
0 .8 1 2
0 .8 4 3
0 .7 9 9
0 .8 0 7
0 .7 9 9
0 .8 1 1
0 .8 1 4
© 2016 John Wiley & Sons Ltd
156 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
T a b le
4 ( C o n ti n u ed )
T es t d a ta
A U C
fo r d iff er en t m o d el s
C o u n tr y
B en ch m a rk
M o d el
1 M o d el
2 M o d el
3 M o d el
4 M o d el
5 M o d el
6 M o d el
7 M o d el
8
S lo v a k ia
(S K )
0 .7 7 7
0 .7 7 4
0 .7 7 6
0 .7 8 0
0 .8 1 1
0 .7 6 9
0 .7 8 6
0 .7 7 7
0 .8 0 8
S lo v en ia
(S I)
0 .7 3 7
0 .7 2 5
0 .7 3 3
0 .6 7 4
0 .7 4 0
0 .7 4 7
0 .7 1 8
0 .7 3 7
0 .7 2 1
S p a in
(E S )
0 .7 3 4
0 .7 1 3
0 .7 3 2
0 .7 0 7
0 .7 9 3
0 .7 2 8
0 .7 5 3
0 .7 3 4
0 .7 7 4
S w ed en
(S E )
0 .8 1 3
0 .8 0 1
0 .8 0 9
0 .7 9 9
0 .7 8 4
0 .8 1 7
0 .8 2 3
0 .8 1 4
0 .8 0 0
U k ra in e (U
A )
0 .7 0 8
0 .7 1 4
0 .7 1 0
0 .7 1 5
0 .7 2 1
0 .7 0 8
0 .7 0 2
0 .7 0 8
0 .7 2 2
U n it ed
K in g d o m
(G B )
0 .6 9 9
0 .7 1 9
0 .6 9 9
0 .6 8 6
0 .7 3 6
0 .6 9 5
0 .7 0 6
0 .6 9 9
0 .7 2 9
U n it ed
S ta te s (U
S )
0 .7 1 0
0 .7 0 1
0 .7 1 1
0 .7 0 9
0 .7 2 2
0 .7 0 5
0 .7 1 6
0 .7 1 0
0 .7 2 3
C o u n tr ie s o n ly
in te st
d a ta
E u ro p ea n
G re ec e (G
R )
0 .7 1 5
0 .6 7 0
0 .7 0 2
0 .7 2 5
0 .7 1 1
0 .7 1 3
0 .7 1 7
0 .7 1 5
0 .7 1 8
L it h u a n ia
(L T )
0 .7 6 7
0 .7 8 2
0 .7 6 7
0 .7 6 4
0 .7 6 8
0 .7 6 9
0 .7 7 5
0 .7 6 7
0 .7 7 8
S er b ia
(R S )
0 .7 3 6
0 .7 1 3
0 .7 3 0
0 .6 0 3 –
0 .8 2 6
0 .7 2 0
0 .7 5 3
0 .7 3 6
0 .7 3 8
U .K
., li q u id a ti o n
d a ta se t
0 .6 0 6
0 .6 2 1
0 .6 0 3
0 .6 2 0
0 .6 1 8
0 .6 1 0
0 .6 0 7
0 .6 0 7
0 .6 3 5
N o n -E u ro p ea n
C h in a (C
N )
0 .5 5 8
0 .5 7 0
0 .5 5 7
0 .5 7 2
0 .5 4 3
0 .5 5 4
0 .5 6 7
0 .5 5 8
0 .5 5 6
C h in a , d el is te d
fi rm
s, D L
0 .5 2 9
0 .5 4 6
0 .5 1 9
0 .5 6 3
0 .7 4 0 + + + +
0 .5 4 2
0 .5 2 0
0 .5 2 9
0 .7 0 7 + + +
S ig n ifi ca n ce : A U C
b et te r th a n b en ch m a rk : 0 .0 0 0 1 = + + + + , 0 .0 0 1 = + + + , 0 .0 1 = + + , 0 .1
= +
A U C
w o rs e th a n b en ch m a rk : 0 .0 0 0 1 = —
, 0 .0 0 1 = — , 0 .0 1 = – , 0 .1
= –
M o d el s:
B en ch m a rk
= T h e L R
m o d el
es ti m a te d
fo r a ll
d a ta
w it h
Z ” -m
o d el
(1 9 8 3 ) v a ri a b le s;
M o d el
1 =
T h e o ri g in a l A lt m a n
(1 9 8 3 ) Z ” -S co re
M o d el ; M o d el
2 =
T h e M D A
m o d el
es ti m a te d fo r a ll d a ta ; M o d el
3 =
T h e L R
m o d el
es ti m a te d fo r a ll d a ta
w it h y ea r d u m m ie s;
M o d el
4 =
T h e
L R
m o d el
es ti m a te d fo r a ll d a ta
w it h si ze
v a ri a b le s;
M o d el
5 =
T h e L R
m o d el
es ti m a te d fo r a ll d a ta
w it h a g e ca te g o ry
d u m m ie s;
M o d el
6 =
T h e
L R
m o d el
es ti m a te d fo r a ll d a ta
w it h in d u st ry
d u m m ie s;
M o d el
7 =
T h e L R
m o d el
es ti m a te d fo r a ll d a ta
w it h co u n tr y ri sk
ra n k in g s;
M o d el
8 =
T h e L R
m o d el
es ti m a te d fo r a ll d a ta
w it h a ll v a ri a b le s.
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 157
Table 5. Differences of Medians Between Non-Failed and Failed Groups
Country WCTA RETA EBITTA BVETD AUC of Z”-Score
Austria (AT) 0.448 0.405 0.126 0.487 0.788 Belgium (BE) 0.223 0.264 0.077 0.431 0.760 Bosnia (BA) 0.106 0.249 0.088 0.580 0.805 Bulgaria (BG) 0.114 0.291 0.094 0.436 0.630 China (CN) 0.068 0.032 0.016 0.280 0.570 China, delisted data, DL 0.089 �0.061 �0.032 0.083 0.546 China, ST data 0.298 0.293 0.139 0.468 0.911 Colombia (CO) 0.226 0.235 0.099 0.705 0.724 Croatia (HR) 0.275 0.335 0.044 0.274 0.812 Czech Republic (CZ) 0.404 0.397 0.069 0.667 0.813 Denmark (DK) 0.267 0.356 0.046 0.733 0.798 Estonia (EE) 0.331 0.388 0.113 1.033 0.827 Finland (FI) 0.402 0.606 0.207 0.907 0.864 France (FR) 0.140 0.213 0.065 0.424 0.723 Germany (DE) 0.131 0.136 0.032 0.262 0.658 Greece (GR) 0.171 0.262 0.049 0.349 0.670 Hungary (HU) 0.175 0.224 0.052 0.587 0.746 Iceland (IS) 0.246 0.261 0.051 0.337 0.674 Ireland (IE) 0.181 0.266 0.039 0.560 0.672 Italy (IT) 0.277 0.164 0.073 0.207 0.799 Latvia (LV) 0.117 0.120 0.042 0.254 0.691 Lithuania (LT) 0.246 0.218 0.051 0.569 0.782 Netherlands (NL) 0.204 0.253 0.077 0.432 0.754 Norway (NO) 0.157 0.219 0.115 0.329 0.694 Poland (PL) 1.340 0.920 0.124 1.351 0.904 Portugal (PT) 0.215 0.200 0.052 0.318 0.724 Romania (RO) 0.222 0.271 0.056 0.298 0.740 Russian Federation (RU) 0.350 0.242 0.069 0.245 0.802 Serbia (RS) 0.120 0.148 0.045 0.389 0.713 Slovakia (SK) 0.256 0.184 0.061 0.431 0.774 Slovenia (SI) 0.111 0.172 0.035 0.326 0.725 Spain (ES) 0.143 0.143 0.076 0.285 0.713 Sweden (SE) 0.255 0.346 0.099 0.663 0.801 Ukraine (UA) 0.204 0.200 0.031 0.449 0.714 United Kingdom (GB) 0.211 0.245 0.033 0.472 0.719 U.K., liquidation dataset 0.156 0.234 0.031 0.447 0.621 United States (US) 0.195 0.378 0.248 0.722 0.701 Average of column items 0.245 0.265 0.073 0.481 0.740 Correlation with Z”-Score AUC
0.611 0.681 0.516 0.574 1.000
WCTA, Working Capital/Total Assets; RETA, Retained Earnings/Total Assets; EBITTA, EBIT/Total Assets; BVETD, Book Value of Equity/Total Liabilities; SALTA, Sales/Total Assets; AUC, Area under the ROC curve; Z”-Score, Altman (1983) Z”-Score in the test data.
© 2016 John Wiley & Sons Ltd
158 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
results (AUC > 0.8) for China (ST firms), Poland, Finland, Estonia, the Czech Republic, Croatia, Bosnia, Russia, and Sweden. Its perfor-
mance is quite low (AUC < 0.7) for Norway, Latvia, Iceland, Ireland, and Germany. The lower part of the table shows the AUCs for the
countries included only in the test data. The performance of the score
is very low in the Chinese CN (primarily private) and DL (delisted)
samples and for liquidation firms in the United Kingdom.
The Appendix shows the medians of the four ratios (X1–X4) by sta- tus and country. Table 5 presents the differences of these medians
between non-failed and failed firms by country. This table also pre-
sents the AUC of the Z”-Score and its correlation with the difference
of medians, which is high for each financial ratio, showing that the
effects of the ratios on AUC are well balanced. For China’s ST firms,
the differences are not exceptionally large, except for EBITTA, which
implies, with the exceptionally high AUC, that ST firms systematically
differ from non-ST firms, although the differences are not extremely
large. The differences between the medians are very large in Poland for
each ratio, justifying the high AUC, and in Finland and the Czech
Republic, where the difference in EBITTA is average. In Germany,
Latvia, China (CN and delisted), and the U.K. (liquidation), the differ-
ences in all four ratios are below average, which is obviously associ-
ated with a low AUC. In the sample of Chinese delisted firms, the
differences in RETA and EBITTA are even negative. In Iceland and
Ireland, the differences only in EBITTA are exceptionally small.
Model 2 in Table 4 is the re-estimated Z”-Score model, where the
coefficients are estimated by MDA for all data. Its AUC (0.745) is
only slightly higher than that for the original model (0.743), supporting
H1 only very weakly, if at all. The classification accuracy in terms of
AR (0.490) is at approximately the average level. The re-estimation of
the coefficients has led to improved classification accuracy in a number
of countries, especially Bosnia, China (ST), Norway, and Greece.
However, it has impaired the classification accuracy in the United
Kingdom and China (delisted). The “Benchmark” column reports the
results for the benchmark model (Z”-Score LR model), showing the
effect of the estimation method. For the benchmark model, the AUC
in all data is 0.748, which is higher than that for Models 1 and 2. The
differences among AUCs are very small, only weakly supporting H2
(estimation method). The LR model (benchmark) and the MDA model
(Model 2) give nearly identical AUCs for each country. This result
was expected because the coefficients of the models are directly compa- © 2016 John Wiley & Sons Ltd
Financial Distress Prediction 159
rable. The similar results for the models may also indicate that the
independent variables conform to multinormality. Nevertheless, these
similarities support what most researchers in the field of default classi-
fication models have concluded: that the accuracy levels of MDA and
logistic regression models are extremely similar.
Model 3 (LR model with year dummies) leads in all test data to a
higher AUC (0.752) than the benchmark model, supporting H3 (bank-
ruptcy year effect). However, the AUC effects are not positive for all
countries. The effects are positive, for example, for Russia, Estonia,
Germany, Ireland, and Latvia, but these effects are not statistically
significant. There are statistically significant negative effects for China
(ST) and Serbia. There are notable negative effects, especially for
Bosnia, Croatia, Hungary, Norway, and Slovenia. These diverse results
are due to the exceptional annual distributions of failed firms in these
countries. For the countries with negative effects, the percent of failed
firms from 2010 (D3 = 1) exceeds 50 per cent, while for the countries with positive effects, it is only a few percent. For each group, non-
failed firms are quite equally distributed over years. When the coeffi-
cient of D3 is very low (�0.666), it strongly decreases the risk estimates of most failed firms in the former countries but of only a few failed
firms in the latter countries, which leads to the observed effects. H3 is
supported by evidence at the level of the whole sample. However, in
some country samples (at the individual country level), H3 is not
supported.
Model 4 (LR model with size variables) performs better than the
benchmark model, which lends support to H4 (size effect) and leads to
AUC = 0.760, indicating AR = 0.520, and to significant improvements in the AUC for, e.g., China (delisted) and Austria. For China
(delisted), the increase in AUC is extremely strong, and the AUC also
improves for, e.g., Estonia, Italy, Slovakia, Spain, and the U.K. Model
4 also led to lower AUCs in a few countries, but this decrease is not
significant. Model 4 is based on the four original variables and the size
effect following a second-order parabola. This kind of size effect is
very small for micro firms but increases when approaching its maxi-
mum value in middle-sized firms. The countries with a positive change
in AUC typically have data in which the percent of failed micro firms
(in all failed firms) is relatively low (40–60 per cent), while there are many middle-sized failed firms. Because the size effect is strongest for
middle-sized firms, the AUC increases. In contrast, the countries with
© 2016 John Wiley & Sons Ltd
160 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
a negative change in AUC have data in which the percentage of failed
micro firms is exceptionally high (70–80 per cent). Model 5 (LR model with age category dummies) gives, for all test
data, nearly the same AUC as the benchmark model (AUC = 0.748). However, the difference is positive and statistically significant, giving
at least marginal support to H5 (age effect). For nearly all countries,
the effect of age on AUC is small. For Austria, however, this effect is
positive and significant. For Austria, the percentage of young firms
(less than 6 years) among non-failed firms is only approximately 10
per cent, whereas this percentage among failed firms is more than 20
per cent. Because Model 5 includes a strong positive age risk effect for
young firms (D2), it increases the risk of many failed firms but of only
a few non-failed firms, which obviously leads to an improvement in
AUC.
Model 6 (LR model with industry dummies) outperforms the bench-
mark model in AUC, supporting H6 (industry effect). It gives an
AUC = 0.751, indicating an AR = 0.502. However, its AUC is notably higher than the benchmark AUC in only a few countries, such as
France, Latvia, Portugal, and Spain. Model 6 has a negative effect on
failure risk for, e.g., restaurants, hotels and the information technology
industry, but it has a positive effect on the construction and manufac-
turing industries. For countries with a positive effect on AUC, the per-
cent of non-failed firms in restaurants, hotels and information
technology is high, while that of failed firms is low. For risky indus-
tries (construction and manufacturing), these distributions are reversed.
Thus, Model 6 gives a positive (negative) risk effect for many failed
(non-failed) firms and a negative (positive) risk effect for only a few
failed (non-failed) firms. Therefore, the AUC increases. This situation
is reversed for the countries with a negative effect on AUC (Austria
and Slovenia). The samples of these countries include a high percent-
age of non-failed firms in the manufacturing industry, leading to a
decrease in AUC.
Model 7 (LR model with the country risk measure) leads to a mar-
ginally higher classification performance (AUC = 0.749) than the benchmark model. This result gives only very weak support for H7
(country of origin effect). However, for each country, the resultant
AUC is nearly identical to that given by the benchmark model. This
result was expected due to the negligible coefficient (�0.003) of the country risk measure (SP country rating rank) in Model 7. This result
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 161
implies that country risk has no effect on the boundary between bank-
rupt and non-bankrupt firms.
Model 8 (LR model with all variables) includes the four financial
ratios and all additional variables and leads to a sizeable increase in
AUC (AUC = 0.771) compared to the benchmark AUC in all test data. However, the effect on the AUCs largely varies and is either neg-
ative or positive in different countries. The positive effect is large in
several countries, such as Estonia, France, Iceland, Italy, Latvia, and
China (delisted). However, it also has a negative effect on AUCs in
several countries, such as Bosnia, Hungary, and Norway. These results
show that the inclusion of additional variables in the original model
will usually increase the AUC, but not in every country.
The “all data” benchmark also performs fairly well for the United
States and Colombian samples (with the U.S.A. firms, unlike the
majority of other firms in this study, being listed or delisted compa-
nies). The poor performance of the predominantly private (CN) and
delisted (DL) Chinese firm samples is associated with very small differ-
ences between the medians of the non-failed and failed groups, as
shown in Table 5. It is clear that the “delisted” status is not compara-
ble with “bankruptcy” status. When the status is defined as “ST”, the
predictability of Chinese listed firms is very high. Prior studies based
on Chinese ST firms have also demonstrated good predictability
(Wang and Campbell, 2010; Zhang et al., 2010). Nevertheless, this
puzzle calls for additional research and modeling work regarding
unlisted and delisted Chinese firms.
6.3. Country-Level Data: Performance of the Z”-Score Models
The heterogeneity of the firms and their distributions in “all data”
make it difficult for a uniform all data model to increase AUCs across
all countries. Table 6 presents the test data AUCs for the different
models estimated for each country separately (country-level models).
In this table, the “all data” Z”-Score LR model acts as the benchmark.
When the models are estimated from country data, this benchmark is
clearly outperformed by the resulting MDA (Model 1) and LR (Model
2) models in only a few countries (Bulgaria, France, Latvia, Spain,
and Sweden). However, these results give only weak support for H1 at
the country level because the effects are not significant. In addition,
the benchmark leads to higher AUCs than in Models 1 and 2, at least
in Austria, Bosnia, Ireland, Slovenia, and the United States. The dif- © 2016 John Wiley & Sons Ltd
162 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
T a b le
6 . T es t D a ta
A U C s fo r D iff er en t E st im
a te d C o u n tr y M o d el s. C o m p a ri so n s a re
W it h th e A U C
o f th e L R
M o d el
E st im
a te d fo r A ll D a ta
( B en ch m a rk )
T es t d a ta
A U C
fo r d iff er en t m o d el s
C o u n tr y
B en ch m a rk
M o d el
1 M o d el
2 M o d el
3 M o d el
4 M o d el
5 M o d el
6 M o d el
7
A u st ri a (A
T )
0 .8 0 0
0 .7 8 2
0 .7 7 0 –
0 .7 7 4
0 .8 2 9
0 .7 8 7
0 .7 0 7 –
0 .7 6 4
B el g iu m
(B E )
0 .7 7 2
0 .7 7 9
0 .7 8 0
0 .8 2 6
0 .7 7 6
0 .7 8 3
0 .7 8 1
0 .8 3 4
B o sn ia
(B A )
0 .8 6 2
0 .8 5 0
0 .8 3 2
0 .8 3 3
0 .8 3 8
0 .8 3 8
0 .8 2 0
0 .8 5 3
B u lg a ri a (B G )
0 .6 8 4
0 .7 8 5
0 .7 8 5
0 .7 8 5
0 .7 9 6
0 .7 9 4
0 .6 7 7
0 .7 0 7
C h in a , S T
co m p a n ie s
0 .9 8 5
0 .9 8 4
0 .9 8 2
0 .9 8 0
0 .9 8 6
0 .9 8 6
0 .9 8 6
0 .9 8 9
C o lo m b ia
(C O )
0 .7 2 6
0 .7 5 5
0 .7 5 4
0 .8 1 8 +
0 .7 8 9
0 .7 4 9
0 .7 7 4
0 .8 2 4 +
C ro a ti a (H
R )
0 .8 4 4
0 .8 6 2
0 .8 5 8
0 .8 6 5
0 .8 5 3
0 .8 6 3
0 .8 4 7
0 .8 6 3
C ze ch
R ep u b li c (C
Z )
0 .8 1 1
0 .8 1 2
0 .8 1 2
0 .8 2 3
0 .8 3 0
0 .8 1 3
0 .8 1 8
0 .8 4 2
D en m a rk
(D K )
0 .8 0 3
0 .8 0 6
0 .8 0 9
0 .8 3 3
0 .8 1 8
0 .8 0 4
0 .8 1 3
0 .8 4 3
E st o n ia
(E E )
0 .8 2 3
0 .8 2 4
0 .8 3 9
0 .8 6 6
0 .8 7 4
0 .8 3 9
0 .8 4 7
0 .8 9 9
F in la n d (F I)
0 .8 6 7
0 .8 6 7
0 .8 7 0
0 .8 8 5
0 .8 6 8
0 .8 7 3
0 .8 7 9
0 .8 9 4
F ra n ce
(F R )
0 .7 3 9
0 .7 7 3
0 .7 9 9
0 .8 0 5
0 .8 2 6
0 .8 0 0
0 .8 1 0
0 .8 4 5
G er m a n y (D
E )
0 .6 7 3
0 .6 8 4
0 .6 8 5
0 .7 3 4
0 .7 0 4
0 .7 0 1
0 .6 9 0
0 .7 6 2
H u n g a ry
(H U )
0 .7 4 2
0 .7 3 6
0 .7 4 3
0 .7 8 8
0 .7 6 7
0 .7 6 5
0 .7 4 7
0 .8 2 0
Ic el a n d (I S )
0 .6 6 4
0 .6 7 0
0 .6 7 1
0 .7 8 4 +
0 .6 8 3
0 .7 1 4
0 .7 0 8
0 .8 2 4 +
Ir el a n d (I E )
0 .6 7 9
0 .6 6 3
0 .6 6 6
0 .7 1 8
0 .6 6 1
0 .6 6 5
0 .6 8 0
0 .7 2 4
It a ly
(I T )
0 .8 0 6
0 .8 0 2
0 .8 1 0
0 .8 3 5
0 .8 3 9
0 .8 0 9
0 .8 1 6
0 .8 6 8
L a tv ia
(L V )
0 .6 7 8
0 .6 9 0
0 .7 0 7
0 .7 4 5
0 .7 2 0
0 .7 1 1
0 .7 3 0
0 .7 7 1
N et h er la n d s (N
L )
0 .7 5 2
0 .7 5 2
0 .7 5 6
0 .7 6 8
0 .7 8 4
0 .7 6 0
0 .7 5 9
0 .8 0 7
N o rw
a y (N
O )
0 .7 1 6
0 .7 4 1
0 .7 4 1
0 .8 0 3
0 .7 4 9
0 .7 4 4
0 .7 4 3
0 .8 1 4
P o la n d (P L )
0 .9 0 3
0 .8 9 1
0 .8 9 7
0 .9 1 5
0 .8 9 9
0 .8 9 6
0 .8 9 5
0 .9 1 3
P o rt u g a l (P T )
0 .7 4 1
0 .7 5 3
0 .7 5 7
0 .7 6 1
0 .7 8 0
0 .7 5 8
0 .7 6 5
0 .7 9 7
R o m a n ia
(R O )
0 .7 5 8
0 .7 1 8
0 .7 7 3
0 .8 0 5
0 .7 5 7
0 .7 5 5
0 .7 5 3
0 .7 7 4
R u ss ia n F ed er a ti o n (R
U )
0 .8 1 1
0 .8 0 8
0 .8 1 0
0 .8 5 4
0 .8 1 1
0 .8 1 1
0 .8 1 8
0 .8 6 2
S lo v a k ia
(S K )
0 .7 7 7
0 .7 7 3
0 .7 7 5
0 .7 8 2
0 .8 1 0
0 .7 6 1
0 .7 7 3
0 .8 0 7
S lo v en ia
(S I)
0 .7 3 7
0 .7 0 3
0 .7 1 9
0 .7 0 4
0 .7 1 1
0 .7 3 9
0 .6 8 6
0 .6 9 8
S p a in
(E S )
0 .7 3 4
0 .7 5 5
0 .7 6 7
0 .7 6 8
0 .8 4 9
0 .7 6 6
0 .7 7 6
0 .8 5 8
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 163
T a b le
6 ( C o n ti n u ed )
T es t d a ta
A U C
fo r d iff er en t m o d el s
C o u n tr y
B en ch m a rk
M o d el
1 M o d el
2 M o d el
3 M o d el
4 M o d el
5 M o d el
6 M o d el
7
S w ed en
(S E )
0 .8 1 3
0 .8 2 6
0 .8 2 9
0 .8 4 3
0 .8 3 1
0 .8 3 1
0 .8 9 2
0 .9 0 1
U k ra in e (U
A )
0 .7 0 8
0 .7 1 2
0 .7 1 2
0 .7 2 0
0 .7 4 2
0 .7 1 8
0 .7 1 8
0 .7 5 8
U n it ed
K in g d o m
(G B )
0 .6 9 9
0 .7 0 8
0 .7 1 3
0 .7 1 8
0 .7 6 4
0 .7 1 4
0 .7 3 1
0 .7 8 5
U n it ed
S ta te s (U
S )
0 .7 1 0
0 .6 8 9
0 .6 8 7
0 .6 8 6
0 .8 1 6 +
0 .6 8 9
0 .7 1 4
0 .8 1 6 +
A v er a g e o f A U C
0 .7 6 8
0 .7 7 3
0 .7 7 8
0 .8 0 1
0 .7 9 9
0 .7 8 2
0 .7 7 9
0 .8 2 3
S ig n ifi ca n ce : A U C
b et te r th a n b en ch m a rk : 0 .0 0 0 1 = + + + + , 0 .0 0 1 = + + + , 0 .0 1 = + + , 0 .1
= + ; A U C
w o rs e th a n b en ch m a rk : 0 .0 0 0 1 = — , 0 .0 0 1 = — ,
0 .0 1 = – , 0 .1
= -.
M o d el s:
B en ch m a rk
= T h e L R
m o d el
es ti m a te d fo r a ll d a ta
w it h Z ” -m
o d el
(1 9 8 3 ) v a ri a b le s;
M o d el
1 = T h e M D A
m o d el
es ti m a te d fo r co u n tr y
d a ta ; M o d el
2 = T h e L R
m o d el
es ti m a te d
fo r co u n tr y d a ta ; M o d el
3 = T h e L R
m o d el
es ti m a te d
fo r co u n tr y d a ta
w it h
y ea r d u m m ie s;
M o d el
4 = T h e L R
m o d el
es ti m a te d fo r co u n tr y d a ta
w it h si ze
v a ri a b le s; M o d el
5 = T h e L R
m o d el
es ti m a te d fo r co u n tr y d a ta
w it h a g e ca te g o ry
d u m m ie s;
M o d el
6 = T h e L R
m o d el
es ti m a te d fo r co u n tr y d a ta
w it h in d u st ry
d u m m ie s;
M o d el
7 =
T h e L R
m o d el
es ti m a te d fo r co u n tr y d a ta
w it h a ll v a ri –
a b le s.
© 2016 John Wiley & Sons Ltd
164 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
ferences in AUCs given by Models 1 and 2 are generally small. Model
1 is clearly outperformed by Model 2 in Romania only. Thus, country-
level evidence does not support H2 (estimation method).
Model 3 (LR model with year dummies) leads in most countries to
a clearly higher AUC than the benchmark model. This evidence gives
support to H3 (bankruptcy year effect). Model 3 leads to a lower
AUC than the benchmark model in Austria, Bosnia, Slovenia, and the
United States. Model 4 (LR model with size variables) leads to
improved performance in nearly every country, supporting H4. This
improvement is significant, however, only in the United States, where
the AUC increases to 0.816 in Model 4, in contrast to the benchmark
model, for which it is only 0.710. This improvement is also notable in
Bulgaria, France, Latvia, Spain, and the United Kingdom. Model 5
(LR model with age category dummies) gives higher AUCs than the
benchmark model for several countries, but none of the improvements
are significant. The positive effect is particularly strong in Bulgaria,
France, and Iceland. Although there are also negative effects on AUC,
this evidence only weakly supports H5 because these negative effects
are relatively small.
Model 6 (LR model with industry dummies) shows both negative
and positive effects on AUCs when compared with the benchmark.
However, the only significant effect is the negative effect found in Aus-
tria. In Bulgaria and Slovenia, the AUC notably decreases due to the
industry dummies. However, there are notable positive effects on the
AUC in, for example, France, Iceland, Latvia, and Sweden. Thus, the
effect is not systematic and gives only weak support to H6. Model 7
(LR model with all variables) leads to a remarkable increase in AUC
compared with the benchmark model in Colombia, Iceland, and the
Unites States and to notable improvements in AUC in Belgium, Den-
mark, Estonia, France, Germany, Hungary, Latvia, Norway, Spain,
and the United Kingdom. Negative effects on AUC are found only in
Austria and Slovenia. Thus, classification performance of the Z”-Score
model with “all test” data remarkably increases in most countries
when different effects are taken into account by additional variables.
This increase is found in most European countries and in Colombia
and the United States. In China, the AUC for the ST sample is extre-
mely high for the Z”-Score LR model, and it can be improved only
slightly by additional variables.
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 165
7. Summary of the Study and a Suggested Extension
The purpose of this study was to assess the classification performance
of the Z”-Score model originally introduced by Altman (1983) using a
very large international dataset. We test how the original version of
the Z”-Score model performs in different countries and how re-estima-
tions using another statistical method and different additional variables
affect the classification performance when the data are very heteroge-
neous. For this kind of testing, seven research hypotheses on classifica-
tion performance are formulated. These hypotheses are tested for all
data and separately for country data (country-level analysis). The
estimation data are from 31 countries, and the results are validated for
34 countries. The countries are mainly from Europe, but three
non-European countries are included (China, Colombia, and the US).
The statuses used in the classification are mainly bankruptcy/active,
but receivership firms are also considered to have failed. In the Chinese
data, ST (special treatment) and delisted firms are also separately
analyzed as failed firms.
The analyses for all data show that the original Z”-Score model per-
forms very satisfactorily in an international context. The effects of the
four financial ratios on performance are well balanced, although Book
Value of Equity/Total Liabilities (BVETD) showed a very small contri-
bution in re-estimation. The original model performs very well in sev-
eral countries, such as Poland, Finland, and China (ST firms). The re-
estimation of the coefficients using MDA only marginally improved
the classification performance, thus weakly supporting the obsolescence
hypothesis (H1) or, put differently, showing that the original coeffi-
cients are extremely robust across countries and over time (opposite
to, e.g., Grice and Ingram, 2001). This same conclusion holds for the
re-estimation of the model using LRA because the performance results
are very similar to those of MDA (H2). The use of additional variables
in the model generally improves the classification accuracy of the origi-
nal model, but the results for countries are dependent on the distribu-
tion of failed and non-failed firms. When the coefficients are estimated
for all data, the effects on performance in a country depend on how
the distributions in that country correspond to the distributions in all
data. For all sets of additional variables, the performance is generally
improved, but the improvement is not strong and the effects vary by
country. Thus, the evidence gives weak support for the effects of all
additional variables. For the effects of bankruptcy year (H3) and size
© 2016 John Wiley & Sons Ltd
166 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
(H4), the effects are stronger, but the variations in the effects between
countries are also stronger. The effects of age (H5), industry (H6), and
country (H7) are marginal. When all additional variables are included
in the same model, the performance generally significantly increases,
but at the same time, the variations among countries become stronger.
In summary, our evidence indicates that the original Z”-Score
model performs well in an international context. It is, however, possi-
ble to extract a more efficient country model for most European coun-
tries and for non-European countries using the four original variables
accompanied by a set of additional background variables. Considering
practical applications, it is obvious that while a general international
model works reasonably well, for most countries, the classification
accuracy may be somewhat improved with country-specific estimation
(a conclusion similar to Xu and Zhang, 2009). In a country model, the
information provided even by simple additional variables may help
boost the classification accuracy to a much higher level.
In finance and accounting research, failure prediction models may
be utilized as risk measures in many different contexts. Where failure
prediction modeling is not the primary focus, it would be time-con-
suming, uneconomical, and superfluous to first estimate a failure pre-
diction model (or models) and then study the phenomenon of interest.
In such instances, a well-tested general model that works reliably and
consistently across different countries is highly desirable. Based on our
empirical tests in this study, the original Z”-Score model and its re-
estimated version, containing the four Altman (1983) study variables
with coefficients re-estimated using a large dataset, work consistently
well internationally and are easy to implement and interpret. Thus, this
kind of accounting-based model can be used by all interested parties,
especially internationally active banks or other financial institutions,
not only for failure or distress prediction but also for other managerial
purposes such as provisioning and economic capital calculation. Inter-
nationally active banks need to develop a universal tool that can be
applied in all subsidiaries and branches to control risk across the
whole banking group.
Further research should focus on other modifications and extensions
than those presented in our paper, such as using alternative modeling
techniques (e.g., panel data analysis), introducing new variables (e.g.,
macroeconomic data), and testing its usefulness with data from other
countries (e.g., emerging markets).
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 167
Notes
1. See Altman and Saunders (1997) for a review of research over this 20 year period. 2. Dichev (1998) compares the Altman Z-score and Ohlson O-score approaches. 3. Grice and Dugan (2003) present a re-estimation of Ohlson’s and Zmijewski’s models. 4. We use bankruptcy, failure, default and financial distress as equivalents. 5. Research devoted to the application of the Z-Score model before 2000 was reviewed by Grice and Ingram (2001). 6. Most of the models focused on stock-exchange-listed firms; thus, the Z’-Score and Z”-Score models were not used. 7. See Zhang et al. (2010) for the rationale for using special-treatment firms as a proxy for bankruptcies. These are firms put on probation by the stock exchange for poor operating performance and/or negative equity. 8. This is done because the results about predictability were also good for such a small sample. 9. These firms are included only in the test data because the predictability of failure was exceptionally poor. 10. From the weighting procedure, it follows that the score (cut-off-value) that best separates failures from non-failures is 0.50 (or, alternatively, 50 per cent). Although the score (logit) in principle has a probability interpretation, the “probabilities” estimated using this weighting scheme in this study do not, however, represent empirical PD’s. It would still require calibration procedures for the models to obtain PD’s that corre- spond to associated empirical PD’s in the population. But this is not attempted in the study, as our focus is more general (the classification accuracies of the models across countries). It is also worth noting that the original Z”-Score does not have a PD inter- pretation either.
References
Acosta-Gonz�alez, E. and F. Fern�andez-Rodr�ıguez, “Forecasting Financial Failure of Firms via Genetic Algorithms,” Computational Economics 43 (2014), pp. 133–157.
Agarwal, V. and R. Taffler, “Comparing the Performance of Market-Based and Accounting-Based Bankruptcy Prediction Models,” Journal of Banking and Finance 32 (2008), pp. 1541–1551.
Altman, E.I., “Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy,” Journal of Finance 23 (1968), pp. 589–609.
Altman, E.I., Corporate Financial Distress: A Complete Guide to Predicting, Avoiding, and Dealing With Bankruptcy (Hoboken: Wiley Interscience, John Wiley and Sons, 1983).
Altman, E.I., “Revisiting the Recidivism – Chapter 22 Phenomenon in the U.S. Bank- ruptcy System,” The Brooklyn Journal of Corporate, Financial & Commercial Law 8 (2014), p. 253.
Altman, E.I. and E. Hotchkiss, Corporate Financial Distress & Bankruptcy, 3rd edn (Hoboken, NJ: John Wiley, 2006).
Altman, E.I. and A. Saunders, “Credit Risk Measurement: Developments Over the Last 20 Years,” Journal of Banking and Finance 21 (1997), pp. 1721–1742.
Balcaen, S. and H. Ooghe, “35 Years of Studies on Business Failure: An Overview of the Classic Statistical Methodologies and Their Related Problems,” The British Accounting Review 38 (2006), pp. 63–93.
© 2016 John Wiley & Sons Ltd
168 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
Bauer, J. and V. Agarwal, “Are Hazard Models Superior to Traditional Bankruptcy Prediction Approaches? A Comprehensive Test,” Journal of Banking and Finance 40 (2014), pp. 432–442.
Beaver, W.H., “Financial Ratios as Predictors of Failure,” Journal of Accounting Research 4 (1966), pp. 71–111.
Beaver, W.H., “Alternative Accounting Measures as Predictors of Failure,” The Accounting Review 43 (1968), pp. 113–122.
Bharath, S.T. and T. Shumway, “Forecasting Default With the Merton Distance to Default Model,” Review of Financial Studies 21 (2008), pp. 1339–1369.
Campbell, J.Y., Hilscher, J. and J. Szilagyi, “In Search of Distress Risk,” Journal of Finance 63 (2008), pp. 2899–2939.
Chava, S. and R.A. Jarrow, “Bankruptcy Prediction With Industry Effects,” Review of Finance 8 (2004), pp. 537–569.
Das, S.R., P. Hanouna and A. Sarin. “Accounting-Based Versus Market-Based Cross- Sectional Models of CDS Spreads,” Journal of Banking and Finance 33 (2009), pp. 719–730.
Dichev, I.D., “Is the Risk of Bankruptcy a Systematic Risk?,” Journal of Finance 53 (1998), pp. 1131–1147.
Dimitras, A.I., Zanakis, S.H. and C. Zopoundis. “A Survey of Business Failures With an Emphasis on Prediction Methods and Industrial Applications,” European Journal of Operational Research 90 (1996), pp. 487–513.
Grice Jr, J.S. and M.T. Dugan, “Re-Estimations of the Zmijewski and Ohlson Bank- ruptcy Prediction Models,” Advances in Accounting 20 (2003), pp. 77–93.
Grice, J.S. and R.W. Ingram, “Tests of the Generalizability of Altman’s Bankruptcy Prediction Model,” Journal of Business Research 54 (2001), pp. 53–61.
Hillegeist, S.A., Keating, E.K., D.P. Cram and K.G. Lundstedt. “Assessing the Proba- bility of Bankruptcy,” Review of Accounting Studies 9 (2004), pp. 5–34.
Hosmer, D. and S. Lemeshow, Applied Logistic Regression (Hoboken: Wiley Inter- science, Prentice Hall International, 1989).
Jackson, R.H.G. and A. Wood, “The Performance of Insolvency Prediction and Credit Risk Models in the UK: A Comparative Study,” British Accounting Review 45 (2013), pp. 183–202.
Kumar, P. and V. Ravi, “Bankruptcy Prediction in Banks and Firms via Statistical and Intelligent Techniques – A Review,” European Journal of Operational Research 180 (2007), pp. 1–28.
Kwak, W., Shi, Y., J.J. Cheh and H. Lee. “Multiple Criteria Linear Programming Data Mining Approach: An Application for Bankruptcy Prediction,” Data Mining and Knowledge Management 3327 (2005), pp. 164–173.
Lyandres, E. and A. Zhdanov, “Investment Opportunities and Bankruptcy Prediction,” Journal of Financial Markets 16 (2013), pp. 439–476.
Merkevicius, E., Gar�sva, G. and S. Girdzijauskas “A Hybrid SOM-Altman Model for Bankruptcy Prediction,” International Conference on Computational Science 3994 (2006), pp. 364–371.
Ohlson, J.A., “Financial Ratios and the Probabilistic Prediction of Bankruptcy,” Jour- nal of Accounting Research 18 (1980), pp. 109–131.
Ooghe, H. and S. Balcaen, “Are Failure Prediction Models Widely Usable? An Empiri- cal Study Using a Belgian Dataset,” Multinational Finance Journal 11 (2007), pp. 33– 76.
Philippe and Partners and Deloitte and Touche, Corporate Finance, Bankruptcy and a Fresh Start: Stigma on Failure and Legal Consequences of Bankruptcy (Brussels, Bel- gium: International Insolvency Institute, 2002).
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 169
Pindado, J., Rodrigues, L. and C. de la Torre. “Estimating Financial Distress Likeli- hood,” Journal of Business Research 61 (2008), pp. 995–1003.
Reisz, A.S. and C. Perlich, “A Market-Based Framework for Bankruptcy Prediction,” Journal of Financial Stability 3 (2007), pp. 85–131.
Ribeiro, S.P., Menghinello, S. and K.D. Backer. ‘The OECD ORBIS Database: Responding to the Need for Firm-Level Micro-Data in the OECD’. No. 2010/1 (Paris: OECD Publishing, 2010).
Shumway, T., “Forecasting Bankruptcy More Accurately: A Simple Hazard Model,” Journal of Business 74 (2001), pp. 101–124.
Smith, M. and D. Liou, “Industrial Sector and Financial Distress,” Managerial Audit- ing Journal 22 (2007), pp. 376–391.
Taffler, R.J., “Empirical Models for the Monitoring of UK Corporations,” Journal of Banking and Finance 8 (1984), pp. 199–227.
Tinoco, M. and N. Wilson, “Financial Distress and Bankruptcy Prediction Among Listed Companies Using Accounting, Market and Macroeconomic Variables,” Inter- national Review of Financial Analysis 30 (2013), pp. 394–419.
Vassalou, M. and Y. Xing, “Default Risk in Equity Returns,” Journal of Finance 59 (2004), pp. 831–868.
Wang, Y. and M. Campbell, “Business Failure Prediction for Publicly Listed Compa- nies in China,” Journal of Business and Management 16 (2010), pp. 75–88.
Wu, Y., Gaunt, C. and S. Gray. “A Comparison of Alternative Bankruptcy Prediction Models,” Journal of Contemporary Accounting and Economics 6 (2010), pp. 34–45.
Xu, M. and C. Zhang, “Bankruptcy Prediction: The Case of Japanese Listed Compa- nies,” Review of Accounting Studies 14 (2009), pp. 534–558.
Zhang, L., Altman, E.I. and J. Yen. “Corporate Financial Distress Diagnosis Model and Application in Credit Rating for Listing Firms in China,” Frontiers of Computer Science in China 4 (2010), pp. 220–236.
Zmijewski, M.E., “Methodological Issues Related to the Estimation of Financial Dis- tress Prediction Models,” Journal of Accounting Research 22 (1984), pp. 59–82.
Appendix. Medians of the Altman (1983) model (Z”) variables by status and country.
Non- failed Failed
Non- failed Failed
Non- failed Failed
Non- failed Failed
Country WCTA RETA EBITTA BVETD
Austria (AT) 0.170 �0.278 0.260 �0.145 0.055 �0.071 0.465 �0.022 Belgium (BE) 0.136 �0.087 0.157 �0.107 0.052 �0.025 0.460 0.029 Bosnia (BA) 0.087 �0.019 0.148 �0.101 0.032 �0.056 0.580 0.000 Bulgaria (BG) 0.216 0.102 0.272 �0.019 0.075 �0.019 0.504 0.068 China (CN) 0.105 0.037 0.064 0.032 0.064 0.048 0.870 0.590 China, delisted data, DL
0.158 0.069 0.286 0.347 0.059 0.091 0.923 0.840
China, ST data 0.106 �0.192 0.281 �0.012 0.052 �0.087 0.776 0.308 Colombia (CO) 0.244 0.018 0.282 0.047 0.104 0.005 0.940 0.235 Croatia (HR) 0.093 �0.182 0.130 �0.205 0.030 �0.014 0.274 0.000
© 2016 John Wiley & Sons Ltd
170 Edward I. Altman, Małgorzata Iwanicz-Drozdowska, Erkki K. Laitinen and Arto Suvas
Appendix (Continued)
Non- failed Failed
Non- failed Failed
Non- failed Failed
Non- failed Failed
Country WCTA RETA EBITTA BVETD
Czech Republic (CZ)
0.196 �0.208 0.230 �0.167 0.050 �0.019 0.591 �0.076
Denmark (DK) 0.128 �0.139 0.306 �0.050 0.005 �0.041 0.758 0.025 Estonia (EE) 0.222 �0.109 0.455 0.067 0.050 �0.063 1.169 0.136 Finland (FI) 0.233 �0.169 0.359 �0.247 0.082 �0.125 0.744 �0.163 France (FR) 0.146 0.006 0.223 0.010 0.059 �0.006 0.501 0.077 Germany (DE) 0.307 0.176 0.150 0.014 0.069 0.037 0.335 0.073 Greece (GR) 0.127 �0.044 0.049 �0.213 0.039 �0.010 0.460 0.111 Hungary (HU) 0.135 �0.040 0.271 0.047 0.054 0.002 0.717 0.130 Iceland (IS) 0.051 �0.195 0.046 �0.215 0.051 0.000 0.196 �0.141 Ireland (IE) 0.198 0.017 0.383 0.117 0.029 �0.010 0.728 0.168 Italy (IT) 0.107 �0.170 0.077 �0.087 0.032 �0.041 0.178 �0.029 Latvia (LV) 0.102 �0.015 0.150 0.030 0.069 0.027 0.342 0.088 Lithuania (LT) 0.205 �0.041 0.273 0.055 0.054 0.003 0.686 0.117 Netherlands (NL)
0.204 0.000 0.303 0.050 0.065 �0.012 0.511 0.079
Norway (NO) 0.185 0.028 0.139 �0.080 0.068 �0.047 0.398 0.069 Poland (PL) 0.198 �1.142 0.258 �0.662 0.076 �0.048 0.850 �0.501 Portugal (PT) 0.164 �0.051 0.133 �0.067 0.028 �0.024 0.343 0.025 Romania (RO) 0.071 �0.151 0.158 �0.113 0.041 �0.015 0.250 �0.048 Russian Federation (RU)
0.088 �0.262 0.106 �0.136 0.043 �0.026 0.194 �0.051
Serbia (RS) 0.093 �0.027 0.129 �0.019 0.027 �0.018 0.389 0.000 Slovakia (SK) 0.137 �0.119 0.185 0.001 0.060 �0.001 0.496 0.065 Slovenia (SI) 0.109 �0.002 0.215 0.043 0.039 0.004 0.434 0.108 Spain (ES) 0.117 �0.026 0.139 �0.004 0.029 �0.047 0.336 0.051 Sweden (SE) 0.266 0.011 0.357 0.011 0.058 �0.041 0.749 0.086 Ukraine (UA) 0.053 �0.151 0.034 �0.166 0.003 �0.028 0.442 �0.007 United Kingdom (GB)
0.179 �0.032 0.294 0.049 0.031 �0.002 0.579 0.107
U.K., liquidation dataset
0.179 0.023 0.294 0.060 0.031 0.000 0.579 0.132
United States (US)
0.164 �0.031 0.374 �0.004 0.003 �0.245 0.800 0.078
Average of medians
a 0.153 �0.092 0.213 �0.050 0.048 �0.025 0.555 0.075
WCTA, Working Capital/Total Assets; RETA, Retained Earnings/Total Assets; EBITTA, EBIT/Total assets; BVETD, Book Value of Equity/Total Liabilities. a Because the two datasets for non-failed U.K. firms are identical, the non-failed medians are used only once in the calculation of the averages.
© 2016 John Wiley & Sons Ltd
Financial Distress Prediction 171