Predicting Private Company Failures in Italy Using Financial and Non-Financial Information

Predicting Private Company Failures in Italy Using Financial
and Non-financial Information
Timothy Wang , The University of Sydney
The study assesses the use of non-financial information in predicting financial distress in private companies
by developing credit risk models tailored to Italian private companies. The in-sample and out-of-sample
prediction test results are indicative of the incremental predictive ability of the two new non-financial variables,
that is, number of shareholders and number of subsidiaries, over accounting ratios and other widely used non-
financial information, including firm age and industry dummies. To be more specific, number of shareholders
and number of subsidiaries are negatively associated with private company failures, and the models augmented
by the two non-financial variables improve forecasting performance from acceptable discrimination to excellent
discrimination over one- to three-year time horizons.
P
rivate companies are the backbone of many research on private company failures has led to several
economies (Berger and Frame 2007). In the EU, calls to develop more robust distress prediction models
for example, small and medium-sized enterprises for private companies (see, e.g., Falkenstein et al. 2000;
(SMEs) represent 99% of all companies, provide approx- Berger and Udell 2006; Beck et al. 2008).
imately 67% of jobs in the private sector and contribute In answer to these calls, this study aims to estimate
to more than 50% of total value-added created by busi- financial distress prediction models that are specifically
nesses (European Parliament 2017). Yet it is very difficult tailored to Italian small businesses in the current eco-
to evaluate the financial conditions of private companies. nomic environment. Italy is chosen because it has expe-
Private companies do not have an active market for their rienced a stagnation in economic growth in recent years
ownership transfer. Credit rating agencies assign credit and private companies in Italy are very vulnerable to
ratings to issuers of debt obligations and debt instru- failure. During the decade prior to the 2007–08 global
ments, which are typically publicly tradeable companies, financial crisis (GFC), Italy’s GDP increased very slowly
governments and sovereign nations rather than private compared with other countries in the EU. Thereafter, the
companies. Since the seminal works of Beaver (1966, GFC had a devastating effect on the Italian economy. In
1968) and Altman (1968), a significant body of research 2012 alone, approximately 1000 businesses failed per day
has modelled corporate financial distress for public com- on average, and GDP declined by 2.5% (Thomas 2013).
panies (e.g., Ohlson 1980; Zmijewski 1984; Constable The International Monetary Fund (2016) has recently
and Woodliff 1994; Jones and Hensher 2004; Hensher forecast that Italy will not fully recover from the GFC
and Jones 2007; Ferguson et al. 2011; Van Peursem and until the mid-2020s.
Chan 2014; Jones et al. 2017). However, it is highly prob- A corporate distress prediction model establishes a
lematic to generalise these models to financial distress relationship between financial distress and accounting
for private companies because of the considerable differ- ratios. While prior studies have demonstrated the
ences between public and private companies (Bhimani predictive capacity of accounting information, other
et al. 2010: 518). Falkenstein et al. (2000), for example, studies have highlighted its limitations. Not only are
demonstrate the different relations of the same account- financial statements prepared under a cost-based system
ing ratios (e.g., cash to total assets, EBIT to interest, total backward-looking, but they also tend to overstate
debts to total assets) to the default of public and private liabilities and understate assets under the influence
companies. of conservatism (Hillegeist et al. 2004). Accounting
Despite their importance and unique features, only a information in private companies is even more opaque
few studies have sought to model financial distress in
private companies over the past 10 years (e.g., Altman
et al. 2010; Bhimani et al. 2010, 2013; Ciampi and Gor-
Correspondence: Timothy Wang, University of Sydney Business
dini 2013; Bhimani et al. 2014; Ciampi 2015; Filipe et al. School, H69, Codrington St, Darlington NSW 2006, Australia. Tel:
2016; Altman et al. 2017), probably because of the greater +61 2 9351 6101; email: [email protected]
difficulty in accessing private datasets. The dearth of Accepted for publication 15 March 2018.
Australian Accounting Review No. 88 Vol. 29 Issue 1 2019 doi: 10.1111/auar.12245 143
18352561, 2019, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/auar.12245 by Kuehne Logistics Universitaet, Wiley Online Library on [07/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Predicting Private Company Failures T. Wang
because there are few disclosure requirements (Ang 1992; group, which is designed to have the benefits of ‘maxi-
Capriat 2009). mizing financial returns, limiting commercial risks and
Beaver et al. (2005), in their prediction of financial dis- expanding its markets’, nevertheless also introduces ‘new
tress for public companies, found that non-accounting complexities to the already complicated corporate sys-
information has incremental predictive ability over ac- tem’ (Wei 2003: 52). Questions then arise as to how
counting ratios. As Table 1 indicates, in the literature number of shareholders and number of subsidiaries are
on private company failures, some studies explore credit associated with the distress of private companies. This
risk for private companies using accounting informa- study extends prior studies by investigating the incre-
tion only (e.g., Edmister 1972; McNamara et al. 1988; mental predictive ability of these two new variables over
Laitinen and Kankaanpaa 1999; Pindado and Rodrigues accounting ratios and other explanatory variables. The
2004; Pompe and Bilderbeek 2005), while other stud- study contributes to the overall literature on corporate
ies control for industry, geography, age and size differ- SME credit risk and on the relative influence of alterna-
ences by matching healthy and distressed private com- tive predictors in particular.
panies by these factors or including them as explanatory
variables (e.g., Bhimani et al. 2010, 2013, 2014; Ciampi
2015; Filipe et al. 2016; Altman et al. 2017). The relevance Methodology
of other non-financial characteristics to private company
failures has been examined in a piecemeal manner: re- Sample selection
porting lags in Keasey and Watson (1987, 1988) and
Peel and Peel (1987); internal credit rating in Grunert This study uses an Italian private company dataset col-
et al. (2005); audit reports in Keasey and Watson (1987); lected from the ORBIS database of Bureau van Dijk.
macroeconomic indicators in Bhimani et al. (2013) and These private companies adopt local generally acceptable
Filipe et al. (2016); and management-related variables accounting principles (GAAP) rather than International
in Keasey and Watson (1987), Bhimani et al. (2013), Financial Reporting Standards (IFRS).1 Using a dataset
Ciampi (2015) and Grunert et al. (2005). from the ORBIS database, Filipe et al. (2016) found that
These non-financial variables are not without prob- data from several European countries, including Italy,
lems. Some macroeconomic indicators do not improve are of better quality with more distress cases and more
predictive accuracy significantly. For example, Filipe complete information than those from Scandinavian
et al. (2016) found that FX rate, unemployment and countries and most Eastern European countries. The col-
economic sentiment indicator, loans granted to non- lection of data requires a definition of active and failed
financial sectors and years to resolve insolvency pro- companies. As Table 1 indicates, prior studies either nar-
ceedings increase the out-of-sample prediction only rowly define financial distress as default of payment or
marginally. Management-related variables are either not failure specific to a country’s legal and financial frame-
readily available or cannot be objectively measured, such works. Consistent with Filipe et al. (2016) and Altman
as management quality in Grunert et al. (2005). More et al. (2017), the present study defines financial distress
importantly, most studies summarised in prior studies very broadly to include a number of failure events be-
mainly fit models to a sample of private firm data with- cause ‘there is no clear framework for SMEs to file for in-
out really understanding how private companies differ solvency’ in Italy (Filipe et al. 2016: 114). Accordingly, an
from public companies. estimation sample of financially distressed private com-
A public company is traded in at least one stock panies is collected for the period 2005–09. It consists of
exchange with shares being purchased and sold by in- companies that experienced: (1) default of payment; (2)
vestors. With a very large number of shareholders and insolvency proceedings; (3) bankruptcy; or (4) in liqui-
prospective investors, a public company has inherent ad- dation. These financial distress events are clearly defined
vantages over a private company, including easy access in the ORBIS database. Default of payment means that
to capital markets and the ability to transfer ownership a debtor has not paid a debt (i.e., loans and/or interest).
easily. In addition, a public company is often a corpo- A firm subject to insolvency proceedings is in admin-
rate group comprised of a parent and its subsidiaries. In istration or receivership or under a scheme of arrange-
contrast, a private company may have only one share- ment. Bankruptcy means that a firm is legally declared
holder or many shareholders. Private companies depend to be unable to pay its debts and so is in the process
on shareholders for inside financing, and shareholders’ of bankruptcy. A firm in liquidation is in the process of
family and friends for the first source of outside financ- selling its assets and will ultimately be dissolved. Given
ing (Ang 1992). Arguably, number of shareholders has a the number of financial distress events, it is not possible
far more significant impact on the survival of a private to give a meaningful ranking in terms of chronological
company given its limited access to capital markets than sequence or severity.
on the survival of a public company. A private company The sample of active firms is drawn from the same
may be a standalone company or a corporate group. A period 2005–09. Active firms are firms that did not
144 Australian Accounting Review

C 2018 CPA Australia
different definitions of distress

T. Wang
C
2018 CPA Australia
Table 1 Applications of financial distress prediction models for private firms

Period Traditional statistical or machine learning
Author/s studied Definition of distress Input variables techniques used
Edmister (1972) 1958–65 Loss borrowers and guarantee recipients Accounting ratios and trends and averages of these Multiple discriminant analysis
ratios only
Keasey and Watson 1970–83 Court winding-up, creditors’ voluntary, Accounting ratios and submission lag, no. of Binary logistic regression
(1987) members’ voluntary and ceased directors, going concern qualification, bank
trading/dissolved floating charge, audit to submission lag, prior and
current years’ qualification
Peel and Peel (1987) 1980–85 Liquidation and entering into receivership Accounting ratios, size and time lag in reporting Binary logistic regression
annual accounts
McNamara et al. 1980–83 Being wound up by court order Accounting ratios only Multiple discriminant analysis
(1988)
Keasey and Watson 1970–80 Liquidation or ceased trading Accounting ratios and delays in submitting financial Multiple discriminant analysis
(1988) statements
Laitinen and 1986–89 Bankruptcy Accounting ratios only Multiple discriminant analysis, binary
Kankaanpaa logistic regression, recursive
(1999) partitioning, the Cox proportional
hazard model and neural networks,
human information processing
Pindado and 1990–93 Sustained non-compliance with banking Accounting ratios only Multiple discriminant analysis and binary
Rodrigues (2004) obligations throughout the whole year logistic regression
Grunert et al. (2005) 1992–96 Loan default Accounting ratios, internal ratings, management Binary probit regression
quality and market position, banks and year
dummy variables
Pompe and 1986–94 Bankruptcy Accounting ratios only Multiple discriminant analysis and neural
Bilderbeek (2005) networks
Bhimani et al. (2010) 1997–2003 Loan default Accounting ratios, size, age, industry and Binary logistic regression
geographical dummies
Bhimani et al. (2013) 1999–2005 Loan default Accounting ratios, GDP, three-month EURIBOR, loan The Cox proportional hazard model
rate, type of management, financial support,
Australian Accounting Review
ownership of assets, management skills, industry

dummies
Predicting Private Company Failures

Ciampi and Gordini Not reported The beginning of formal legal proceedings 10 accounting ratios only Neural networks, multiple discriminant
(2013) for debt (bankruptcy, forced liquidation analysis and logistic regression
etc.)
Bhimani et al. (2014) 1996–2003 Loan default Accounting ratios, size, age, business volatility, Binary mixed logistic regression
industry and geographical dummies
Ciampi (2015) 2008–10 The beginning of formal legal proceedings Accounting ratios, inside director ownership, outside Binary logistic regression
for debt recovery (bankruptcy, forced director ownership, venture capital ownership,
liquidation, etc.) size, industry and geographical dummies
(Continued)
145
experience any of the above financial distress events.
Traditional statistical or machine learning
Multiple discriminant analysis and binary

Firms of other status (i.e., unknown situations, dormant,
Multi-period logit model and the Cox

reorganisation, mergers or takeovers) are excluded from
active firms to avoid ambiguity. An attempt was made
techniques used
proportional hazard model

to collect up to three annual reporting periods’ data
on financially distressed and active companies. For fi-
nancially distressed companies, observations up to three
logistic regression
years prior to any of the above financial distress events
were selected, and for active companies, observations
were randomly selected from the available firm-years.
Following Ohlson (1980) and Jones and Hensher (2004),
the dataset used in this study includes newly founded
companies and companies having one or two years’ fi-
nancial statement records. To avoid the under-sampling
of active firms associated with a matched sample, the
Accounting ratios, size, year and industry dummies,
economic sentiment indicator, loans granted to
non-financial sector, years to resolve insolvency
study uses an unbalanced sample, which better reflects

proceedings, shareholders, geographical and
the actual failure rates. The performance of a logit model

Accounting ratios, FX rate, unemployment,
is very robust regardless of whether the model is esti-

mated on a matched or unbalanced sample (Crone and
Finlay 2012). At the time of data collection, a data filter
Input variables
was used to exclude observations with no or very lim-

ited financial statement records. The estimation sample
SP country rating rank
consists of 523 160 firm-years.

To validate the out-of-sample prediction accuracy of
industry dummies
an estimated model, a test sample is also developed. A

test sample from the same period as the original sample is
not useful for intertemporal validation, thereby provid-
ing an inadequate test of a model’s predictive accuracy
(Joy and Tollefson 1975; Charitou et al. 2004; Jones and
Hensher 2004). Hence, a test sample is collected from
another period, namely, 2010–13, using the same defini-
tions and procedures applied to the estimation sample.
disappearing from the sample before
special cases depending on countries

2010 with negative equity in the last
Defaulted, in receivership, bankrupt, in
This produces a test sample of 645 752 firm-years.

Bankruptcy, or under receivership, or
liquidation, no updated status
Definition of distress
(e.g., liquidation in Norway)
Explanatory variables
The study includes number of subsidiaries and number

of shareholders as two new explanatory variables for fi-
nancial distress in private companies in order to assess
their incremental predictive content. To select other pre-
year
dictors, a number of financial ratios used in prior studies

were extracted. These financial ratios cover a firm’s liq-
uidity, solvency, profitability, activity and coverage. In
the literature, there is no consensus on the selection of
studied
accounting ratios (Balcaen and Ooghe 2006: 84). Fol-

Period
2000–09
2007–10
lowing the approach by Filipe et al. (2016), accounting

ratios were selected based on their associations with the
dependent variable and their popularity in prior studies.
After examining the association of one accounting ratio
Table 1 Continued
Altman et al. (2017)
with the dependent variable at a time in univariate analy-

Filipe et al. (2016)
ses, those accounting ratios with an area under the curve

(AUC) of the Receiver Operating Characteristic (ROC)
above 0.65 were retained. When two accounting ratios
Author/s
are highly correlated with the correlation coefficient be-

ing 0.6 or above, the ratio with a larger AUC was retained.

T. Wang Predicting Private Company Failures
Table 2 Sample description failure rate during the 2005–13 period was 6.55%, the
Panel A: Failure rates by year
annual failure rates show considerable fluctuations and
generally reflect the overall economic health of Italy, with
Year Total Failed % higher annual failure rates during the GFC and lower
2005 10 144 6312 62.22 failure rates following the GFC. The number of total
2006 148 970 7597 5.10 observations is very small in 2005, 2009 and 2013, and
2007 178 034 9156 5.14 so the percentages of failed observations may not repre-
2008 186 012 9509 5.11 sent the actual annual failure rates for these years. Panel
2009 8840 8839 99.99
2010 208 126 9582 4.60
B shows the distribution of failures by industry break-
2011 217 406 12 616 5.80 down. The highest average failure rate is in public admin-
2012 207 778 9450 4.55 istration (13.48%), followed by construction (8.38%),
2013 3602 3490 96.89 manufacturing (7.39%), retail trade (6.99%), wholesale
Total 1 168 912 76 551 6.55 trade (6.98%), and transportation, communication and
public utilities (6.15%). The lowest average failure rate
Panel B: Failure rates by industry
is in agriculture, forestry and fishing (2.62%), followed
Industry Total Failed % by finance and real estate (4.23%), mining (4.89%) and
services (5.51%). The variations in average failure rates
Agriculture, forestry & 15 887 416 2.62
fishing across industries indicate the need to control for industry
Mining 4134 202 4.89 influences in financial distress prediction models.
Construction 176 779 14 806 8.38 A correlation matrix of all the independent variables
Manufacturing 279 809 20 664 7.39 used in the study is shown in Table 3. All the correlations
Transportation, 70 534 4339 6.15
are statistically significant at the 1% level. The highest
communication and
public utilities correlation of –0.546 is between working capital to total
Wholesale trade 186 184 12 991 6.98 assets and fixed assets to total assets, followed by a corre-
Retail trade 82 077 5737 6.99 lation of –0.331 between working capital to total assets
Finance and real estate 163 930 6935 4.23 and total liabilities to total assets. These correlations are
Services 189 311 10 425 5.51
not large enough to cause any serious concerns about
Public administration 267 36 13.48
Total 1 168 912 76 551 6.55 multicollinearity and are broadly consistent with what
has been reported in previous studies (see, e.g., Altman
This table provides summary statistics for the sample. In Panel 1968; Ohlson 1980).
A, the annual failure rates show considerable fluctuations during
Table 4 presents the mean values of the firm-specific
the 2005–13 period. Panel B provides failure rates by industry
breakdown, suggesting that the failure rates vary considerably independent variables for the first to third years prior to
across industries. failure. First, it can be observed that distressed firms
differ considerably from active firms in terms of the
Following Charitou et al. (2004) and Filipe et al. (2016), mean values of these independent variables, suggesting
the study applies a forward selection procedure to find that distressed firms have fewer shareholders and sub-
the best combination of the retained ratios in multi- sidiaries, are younger, less liquid, less profitable and more
variate analyses. The procedure results in these ratios as highly leveraged. Second, the mean values of all the ratios
explanatory variables: cash to total assets; EBIT to total for failed firms show a deteriorating trend as the year of
assets; fixed assets to total assets; total liabilities to total failure approaches, whereas the mean values of the active
assets; and working capital to total assets. The study also firms show a decreasing trend for cash to total assets and
includes firm age and industry dummy variables because EBIT to total assets, and an improving trend for fixed
these control variables are widely used in the literature.2 assets to total assets, total liabilities to total assets and
Following Shumway (2001), the firm-specific variables working capital to total assets. The summary statistics
were winsorised at the 99th and 1st percentiles to reduce provide some preliminary indication of the predictive
the effect of outliers on statistical results. capacity of the selected explanatory variables prior to
failure.
Empirical results
Estimation procedure
Descriptive statistics
This study models financial distress in two states: zero
Table 2 provides summary statistics for the whole sam- for active and one for failed. Hence, a statistical tech-
ple. Panel A shows that of 1 168 912 accounts, 76 551 nique needs to allow for a binary dependent variable.
relate to accounts that experienced financial distress, rep- Jones et al. (2015) examined the predictive ability of 20
resenting 6.55% of the total. While the average annual binary classifiers in predicting credit ratings changes. At

C 2018 CPA Australia Australian Accounting Review 147
148
Table 3 Pearson product-moment correlations between significant covariates for the entire sample
Firm age CA/TA EBIT/TA FA/TA TL/TA WC/TA No. of subs No. of sh AFF MI CO MA TCP WT RT FIR SR PA

Firm age 1.000
CA/TA −.074 1.000
EBIT/TA −.005 .194 1.000
FA/TA .150 −.231 −.156 1.000
TL/TA −.264 −.180 −.220 −.280 1.000
WC/TA .058 .230 .238 −.546 −.331 1.000
No. of subs .064 −.037 −.021 .087 −.055 −.030 1.000
No. of sh .068 −.009 .005 .051 −.063 .008 .027 1.000
AFF .008 −.037 −.042 .119 −.030 −.056 −.009 .005 1.000
MI .043 −.013 −.003 .025 −.025 −.007 .008 .005 −.007 1.000
CO −.072 −.079 −.015 −.146 .096 .130 .003 −.012 −.050 −.025 1.000
MA .128 −.008 .042 −.064 −.018 .012 −.029 −.007 −.066 −.033 −.237 1.000
TCP −.032 .019 −.006 .022 .024 −.061 .005 .003 −.030 −.015 −.107 −.142 1.000
WT −.010 .040 .041 −.202 .036 .044 −.043 −.020 −.051 −.026 −.184 −.244 −.110 1.000
RT −.039 .109 −.010 −.060 .051 −.026 −.019 −.008 −.032 −.016 −.116 −.154 −.070 −.120 1.000
FIR .038 −.110 −.084 .291 −.151 .004 .031 .017 −.047 −.024 −.170 −.227 −.102 −.176 −.111 1.000
SR −.066 .077 .028 .128 −.003 −.111 .055 .024 −.052 −.026 −.186 −.247 −.111 −.191 −.121 −.178 1.000
PA −.008 .001 −.003 .000 .004 .001 .000 −.002 −.002 −.001 −.006 −.008 −.004 −.007 −.004 −.006 −.007 1.000
Firm age (in years); CA/TA = cash/total assets; EBIT/TA = EBIT/total assets; FA/TA = fixed assets/total assets; TL/TA = total liabilities/total assets; WC/TA = working capital/total assets; AFF = one for
agriculture, forestry and fishing, zero otherwise; MI = one for mining, zero otherwise; CO = one for construction, zero otherwise; MA = one for manufacturing, zero otherwise; TCP = one for
transportation, communications and public utilities, zero otherwise; WT = one for wholesale trade, zero otherwise; RT = one for retail trade, zero otherwise; FIR = one for finance, insurance and
real estate, zero otherwise; SR = one for services, zero otherwise; PA = one for public administration, zero otherwise. All correlations are significant at the 0.01 level (two-tailed).
C

2018 CPA Australia
T. Wang
Table 4 Descriptive statistics for firm-specific variables

Fixed assets Total Working
Years prior Distress Cash to EBIT to total to total liabilities to capital to Number of Number of
to failure state Statistics Firm age total assets assets assets total assets total assets subsidiaries shareholders
One Active Mean 15.04 0.08 0.04 0.33 0.74 0.15 0.60 2.85
Std. dev. 13.04 0.13 0.09 0.31 0.26 0.32 2.18 5.73
Failed Mean 12.26 0.05 −0.04 0.25 0.96 −0.04 0.20 1.15
Std. dev. 11.98 0.11 0.16 0.26 0.24 0.38 0.86 0.98
Two Active Mean 14.82 0.09 0.05 0.32 0.76 0.14 0.62 2.89
Std. dev. 12.93 0.13 0.09 0.31 0.25 0.32 2.22 5.86
Failed Mean 12.17 0.05 −0.01 0.25 0.92 0.01 0.22 1.16
Std. dev. 11.96 0.11 0.13 0.26 0.21 0.35 0.89 0.98
Three Active Mean 15.03 0.09 0.05 0.31 0.76 0.14 0.63 2.93
Std. dev. 12.80 0.13 0.09 0.30 0.24 0.30 2.32 5.98
Failed Mean 12.05 0.06 0.01 0.25 0.89 0.05 0.24 1.18
Std. dev. 11.98 0.11 0.11 0.26 0.20 0.32 0.94 1.04
This table provides the mean values of the firm-specific variables of failed and active firms over one- to three-year time horizons. Overall,
failed firms differ considerably from active firms in terms of the mean values of the firm-specific variables. On average, failed firms have
fewer shareholders and subsidiaries. With the year of failure approaching, the mean values of all the ratios show a deteriorating trend
among failed firms; by contrast, the mean values of most ratios show an improving trend over time among active firms. The mean values
of number of subsidiaries and number of shareholders remain relatively stable for both failed and active firms.
one extreme are traditional statistical techniques, such independent variable on the outcome, one can report
as multiple discriminant analysis, logistic regression and the odds ratio. For a continuous independent variable,
probit regression, which produce highly interpretable the estimated odds ratio approximates how much more
outputs. At another extreme are machine learning tech- likely or unlikely the risk of the outcome increases for
niques such as neural networks, support vector ma- every increase of one unit in the continuous indepen-
chines and generalised boosting models, which gener- dent variable. However, such a statement is questionable
ate black-box outputs. Although these machine learning because the additional risk of the outcome depends on
techniques are better at capturing nonlinear relation- the level of the continuous independent variable. We are
ships and interactions in a dataset, Jones et al. (2015) faced with the dilemma when a continuous variable is
found that traditional statistical techniques such as logit modelled linearly in the logit (Hosmer and Lemeshow
and probit compared rather well with machine learning 2000: 64).
techniques in terms of predictive accuracy. A logistic re- A solution to the dilemma is to explore the effect of
gression model is Pr(Y = 1|X ) = (β0 + βX ), where Y an independent variable on the outcome by its marginal
is a binary outcome variable (e.g., one for failed and zero effect. The marginal effect measures the expected in-
for active), X is a vector of independent variables, β0 and stantaneous change in the event probability for a change
β are estimated regression coefficients for constant and in an independent variable when all other independent
independent variables and is the cumulative standard variables are held constant. The marginal effect depends
logistic distribution function. A probit regression model on the values of all independent variables and their es-
takes the form Pr(Y = 1|X ) = (β0 + βX ), where is timated coefficients; therefore, one has to decide on the
the cumulative standard normal distribution. As Table 1 levels of the independent variables in the computation
indicates, logit has gained more popularity than pro- of the marginal effect. One approach is to compute the
bit although prior studies found that probit and logit marginal effect at every observation and then calculate
yielded similar results (e.g., Jones et al. 2015, 2017). the sample average.
Given its predictive accuracy, high generalisability, su- This study estimates logistic regression models on the
perior interpretability and popularity, this study uses a estimation sample in predicting financial distress one to
standard logit approach to model financial distress. three years ahead based on the explanatory variables de-
As pointed out by Train (2009), the logistic choice scribed above. The models estimated are single-period
probability function has several desirable characteris- models because each model uses only one observation
tics. Firstly, P ( y i = 1|x i ) is confined to values between for each firm in an estimation sample assuming con-
zero and one. Secondly, the choice probabilities for all secutive annual accounts are independent. These single-
alternatives add up to one. Thirdly, the sigmoid or S- period models are estimated using maximum likelihood.
shaped relationship between the logistic choice prob- To assess the goodness-of-fit of these models, the study
ability and x i can better reflect the non-constant ef- uses various measures, including the in-sample and out-
fect of one unit increase in an independent variable on of-sample classification of firms and AUCs. When clas-
the estimated probability. To examine the impact of an sifying firms based on a single cut-off point, one can

make Type I and II errors, the errors of classifying failed Model I achieves an AUC of 0.763 in the in-sample re-
firms as active and active firms as failed respectively. sults. This is increased only marginally to AUCs of 0.767
Classification based on a single cut-off point, however, and 0.766 for Models II to III respectively, but very sig-
provides a very limited view of the predictive capacity of nificantly to AUCs of 0.938 and 0.802 for Models IV and
a model. An alternative measure is the area under a ROC V respectively. In contrast, the out-of-sample prediction
curve that plots the true positive rate (sensitivity) against of Model I is increased only marginally by Models II
the false positive rate (1-specificity) at various cut-off and V, but significantly by Model IV. Given these re-
points. The ROC is a commonly used measure in the sults, industry and firm age improve performance only
literature for comparing the performance of classifiers marginally and number of shareholders and number of
(Swets et al. 2000; Jones et al. 2015). The AUC represents subsidiaries – number of shareholders in particular –
the probability that a randomly selected failed firm is improve performance very significantly.
rated with greater suspicion of failure than a randomly After assessing the incremental predictive ability of
selected active firm. A model with an AUC value of 0.5 non-accounting information, Table 6 presents the full
has no discrimination, 0.7 to 0.8 acceptable discrimina- logistic regression models for the first, second and third
tion, 0.8 to 0.9 excellent discrimination and 0.9 or above year prior to failure for the 2005–09 estimation sample
outstanding discrimination (Hosmer and Lemeshow incorporating all the accounting and non-accounting
2000: 162). information. While Panel A reports the coefficient es-
timates, test statistics for the null hypothesis that each
coefficient is equal to zero, and marginal effect, Panels B
Parameter estimates and C provide the in-sample and out-of-sample predic-
tion test results respectively.
To provide evidence of the incremental predictive ability Panel A of Table 6 reports the likelihood ratio tests for
of non-accounting information over accounting infor- the overall significance of the coefficients for the selected
mation, particularly number of shareholders and num- explanatory variables in the logistic analysis for the first,
ber of subsidiaries, five models are estimated for the first second and third year prior to failure. Panel A reports
year prior to failure. Model I includes only accounting the likelihood ratio test values of 35 259, 36 880 and 35
ratios, Model II accounting ratios and industry dum- 399 for the first, second and third year prior to failure
mies, Model III accounting ratios and firm age, Model models respectively. The p-value for each likelihood ratio
IV accounting ratios and number of shareholders, and test is P [χ2 (17)] = 0.000, which is significant at the 1%
Model V accounting ratios and number of subsidiaries. level. The null hypothesis is rejected and at least one and
Panel A of Table 5 presents the estimated coefficients, perhaps all the coefficients are different from zero. In
and z-values for the five model specifications. One can addition to the likelihood ratio tests, Panel A reports
conduct a likelihood ratio test to evaluate the incremen- the univariate Wald test statistics. Under the hypothesis
tal information content of each of the non-accounting that a coefficient is zero, these statistics follow a standard
variables. The likelihood ratio test is based on the statistic normal distribution. It can be clearly observed that most
G = Deviance (model without explanatory variables) less independent variables are significant at the 1% level over
Deviance (model with explanatory variables). The like- one- to three-year time horizons.
lihood ratio statistic, the difference between the residual Firm age, number of shareholders, number of sub-
deviance of the model with industry dummies (Model sidiaries, cash to total assets, EBIT to total assets, fixed
II) and without industry dummies (Model I) is 383. assets to total assets and working capital to total assets
With nine degrees of freedom, the chi-square statistic is are inversely correlated with the probability of private
statistically significant at the 1% level, suggesting the in- company failure. In contrast, total liabilities to total as-
cremental information content of industry. Similarly, the sets are positively related to the probability of private
likelihood ratio statistics for firm age, number of share- company failure. Compared with firms in agriculture,
holders and number of subsidiaries are 301, 24 135 and fishing and forestry, firms in other industries are more
3286 respectively. With one degree of freedom, the chi- likely to enter into financial distress. The marginal effect
square statistics are statistically significant at the 1% level. is used to examine how much the probability of company
The study conducts similar likelihood ratio tests over failure changes when an explanatory variable is changed
two- to three-year time horizons and provides evidence by one unit. The marginal effect is computed at each
of the incremental predictive value of non-accounting observation and then the sample average of individual
information again although the results are not reported marginal effects for the overall marginal effect of every
here. independent variable is calculated.
In order to further evaluate the incremental predictive Over a one-year time horizon, an increase of one year
ability of number of shareholders and number of sub- in firm age decreases the probability of financial distress
sidiaries, Panels B and C of Table 5 report the in-sample by 0.001%. The marginal effect for number of share-
and out-of-sample predictive accuracy of Models I to V. holders is −0.023%, suggesting that an increase of one

Table 5 Incremental information content of non-accounting information over a one-year time horizon

T. Wang
C
Panel A: Estimation results

2018 CPA Australia
Model I Model II Model III Model IV Model V

Independent
variables Estimates z-value Estimates z-value Estimates z-value Estimates z-value Estimates z-value
Accounting ratios
CA to TA −2.797∗∗∗ (−22.868) −2.663∗∗∗ (−21.452) −2.962∗∗∗ (−24.040) −2.846∗∗∗ (−22.235) −3.011∗∗∗ (−24.375)
EBIT to TA −0.866∗∗∗ (−7.191) −0.955∗∗∗ (−7.859) −0.984∗∗∗ (−8.257) −0.220∗ (−1.811) −0.871∗∗∗ (−7.294)
FA to TA −2.220∗∗∗ (−37.658) −2.047∗∗∗ (−32.743) −2.163∗∗∗ (−37.070) −2.008∗∗∗ (−32.579) −2.012∗∗∗ (−34.243)
TA to TL 2.537∗∗∗ (28.345) 2.664∗∗∗ (28.882) 2.261∗∗∗ (25.130) 1.840∗∗∗ (21.084) 2.270∗∗∗ (25.779)
WC to TA −1.657∗∗∗ (−32.328) −1.600∗∗∗ (−29.924) −1.622∗∗∗ (−32.147) −1.322∗∗∗ (−24.877) −1.658∗∗∗ (−32.431)
Industry dummies
Mining 0.486∗∗ (1.965)
Construction 0.646∗∗∗ (4.818)
Manufacturing 0.760∗∗∗ (5.713)
Transportation 0.477∗∗∗ (3.451)
Wholesale trade 0.437∗∗∗ (3.247)
Retail trade 0.454∗∗∗ (3.302)
Finance and real 0.127 (0.928)
estate
Services 0.283∗∗ (2.103)
Public 1.893∗∗∗ (5.367)
administration
Non-accounting variables
Firm age −0.019∗∗∗ (−16.600)
No. of −4.298∗∗∗ (−63.457)
shareholders
No. of subsidiaries −1.429∗∗∗ (−38.385)
Constant −4.116∗∗∗ (−44.868) −4.791∗∗∗ (−29.412) −3.669∗∗∗ (−38.907) 2.166∗∗∗ (18.921) −3.636∗∗∗ (−40.163)
Null deviance 79 587 79 587 79 587 79 587 79 587
Residual deviance 70 920 70 537 70 619 46 785 67 634
Likelihood ratio 8667∗∗∗ 9050∗∗∗ 8968∗∗∗ 32 802∗∗∗ 11 953∗∗∗
test
Pseudo R square 0.130 0.136 0.134 0.461 0.178

Panel B: In-sample prediction tests
Model I Model II Model III Model IV Model V

192 749 (182 587 active firms and 10 162 failed firms)
Classification 70.2% (128 084 active firms 70.4% (128 450 active firms 70.4% (128 481 active firms 87.6% (159 976 active firms 73.3% (133 828 active firms
accuracy and 7137 failed firms and 7156 failed firms and 7124 failed firms and 8906 failed firms and 7453 failed firms
correctly classified using the correctly classified using the correctly classified using the correctly classified using the correctly classified using the
optimal cut-off point of 006) optimal cut-off point of 0.06) optimal cut-off point of 0.06) optimal cut-off point of 0.14) optimal cut-off point of 0.07)
AUC 0.763 0.767 0.766 0.938 0.802
151
(Continued)
This table summarises the results of the model controlling for accounting information only (Model I), the model controlling for accounting information and industry dummies (Model II), the model
information and number of subsidiaries (Model V) for the first year prior to failure in order to assess the relative influence of industry, firm age, number of shareholders and number of subsidiaries
significance at the 1%, 5% and 10% levels respectively. Overall the study yields evidence of the incremental predictive ability of each of these non-accounting variables over accounting ratios. Panels
B and C display in-sample and out-of-sample predictive accuracy in terms of classification of firms and the area under the ROC curve. The ROC curve plots the true positive rate against the false
controlling for accounting information and firm age (Model III), the model controlling for accounting information and number of shareholders (Model IV), and the model controlling for accounting
optimal cut-off point of 0.06)
separately. CA/TA = cash/total assets; EBIT/TA = EBIT/total assets; FA/TA = fixed assets/total assets; TL/TA = total liabilities/total assets; WC/TA = working capital/total assets. ∗∗∗ , ∗∗ and ∗ denote
shareholder reduces the probability of failure by 0.023%.
correctly classified using the

70.4% (153 016 active firms
In contrast, an increase of one subsidiary reduces the
and 10 764 failed firms

probability of failure by only 0.007%. As for account-
ing ratios, an increase of one unit in cash to total as-
Model V
0.778
sets, EBIT to total assets, fixed assets to total assets and
working capital to total assets decreases the probability
of private companies entering into distress by 0.015%,
0.002%, 0.009% and 0.006% respectively. In contrast,
the marginal effect for the ratio of total liabilities to total
assets is 0.010%, suggesting that an increase of one unit
in the ratio increases the probability of private company
failure by 0.010%. The marginal effects of the industry

dummies are positive. Over a one-year time horizon,

private companies in public administration are 0.014%
Model IV
0.897
times more likely to fail than those in agriculture, fishing

and forestry. One can interpret the marginal effects of
other industry dummies in a similar way.
Goodness-of-fit
232 652 (217 443 active firms and 15 209 failed firms)

This section explores how effectively the logistic regres-

sion models augmented by number of shareholders and

number of subsidiaries describe the outcome variable
Model III
0.767
over one- to three-year time horizons, which is often re-

ferred to as goodness-of-fit. Panel A of Table 6 reports
the pseudo R indices over one- to three-year time hori-
zons, which are 0.492, 0.493 and 0.509 respectively. This
measure, however, is very difficult to interpret in logistic
analysis: a logistic regression model is not estimated to
70% (152 139 active firms and
minimise variance and the OLS approach to goodness-

10 653 failed firms correctly
classified using the optimal
of-fit is not applicable. To better evaluate the predictive

power of these logistic regression models, this study val-
cut-off point of 0.06)
idates these models on the test sample. The test sample

Model II
0.774
over one- to three-year time horizons consists of 232

652, 219 335 and 193 765 firms respectively.
The model for the first year prior to failure classifies
88.6% of firms in the estimation sample based on the
optimal cut-off point of 0.13. While prior studies have
developed models on matched samples and classified
firms based on the cut-off score of 0.5, this study uses

much lower cut-off scores because the models are esti-

mated on the unbalanced sample. This study also uses

AUCs to assess the predictive ability of the model for the
Model I
0.772
first year prior to failure. The model achieves an AUC of

Panel C: Out-of-sample prediction tests
0.946 and thereby provides outstanding in-sample dis-

positive rate at various cut-off scores.
crimination. The logistic analysis achieves similar one-

year-ahead out-of-sample prediction accuracy: 84.2% of
the firms correctly classified based on the optimal cut-off
score of 0.05 and an AUC of 0.894. In unreported results,
Table 5 Continued
the reduced model (i.e., the model without controlling

for number of shareholders and number of subsidiaries)
Classification
achieves out-of-sample classification of 69.5% and an

accuracy
AUC of 0.769.
As the time horizon for private company failures in-
AUC
creases, the predictive capacity of the models declines.

Table 6 Logit parameter estimates, marginal effects and in-sample and out-of-sample prediction tests: Years 1 to 3
C

Panel A: Estimation results
T. Wang
One year prior to failure Two years prior to failure Three years prior to failure
Marginal Marginal Marginal

Independent variables Estimates z-value effect (%) Estimates z-value effect (%) Estimates z-value effect (%)
2018 CPA Australia

Non-accounting variables
Age −0.004∗∗∗ −3.758 −0.001∗∗∗ −0.002∗∗ −2.192 −0.000∗∗ −0.001 −0.942 −0.000
No. of shareholders −4.274∗∗∗ −62.963 −0.023∗∗∗ −3.768∗∗∗ −76.889 −0.068∗∗∗ −3.704∗∗∗ −81.349 −0.107∗∗∗
No. of subsidiaries −1.210∗∗∗ −31.612 −0.007∗∗∗ −1.067∗∗∗ −35.536 −0.019∗∗∗ −0.942∗∗∗ −33.195 −0.027∗∗∗
Accounting variables
Cash to total assets −2.809∗∗∗ −21.268 −0.015∗∗∗ −2.760∗∗∗ −22.538 −0.050∗∗∗ −2.534∗∗∗ −19.917 −0.073∗∗∗
EBIT to total assets −0.362∗∗∗ −2.919 −0.002∗∗∗ −1.633∗∗∗ −13.637 −0.030∗∗∗ −2.305∗∗∗ −17.017 −0.067∗∗∗
Fixed assets to total assets −1.586∗∗∗ −23.745 −0.009∗∗∗ −1.280∗∗∗ −20.044 −0.023∗∗∗ −0.907∗∗∗ −13.209 −0.026∗∗∗
Total liabilities to total assets 1.808∗∗∗ 19.860 0.010∗∗∗ 1.482∗∗∗ 17.558 0.027∗∗∗ 1.281∗∗∗ 14.476 0.037∗∗∗
Working capital to total −1.182 ∗∗∗ −20.857 −0.006∗∗∗ −0.844∗∗∗ −15.523 −0.015∗∗∗ −0.668∗∗∗ −11.087 −0.019∗∗∗
assets
Industry dummies
Mining 0.836∗∗∗ 3.002 0.007∗∗ 1.219∗∗∗ 4.847 0.043∗∗∗ 1.024∗∗∗ 3.818 0.051∗∗
Construction 0.762∗∗∗ 5.353 0.005∗∗∗ 0.937∗∗∗ 6.622 0.024∗∗∗ 1.001∗∗∗ 6.701 0.043∗∗∗
Manufacturing 1.009∗∗∗ 7.146 0.007∗∗∗ 1.156∗∗∗ 8.246 0.030∗∗∗ 1.040∗∗∗ 7.033 0.040∗∗∗
Transportation 0.499∗∗∗ 3.399 0.003∗∗∗ 0.637∗∗∗ 4.360 0.016∗∗∗ 0.618∗∗∗ 4.017 0.024∗∗∗
Wholesale trade 0.582∗∗∗ 4.075 0.004∗∗∗ 0.710∗∗∗ 5.003 0.017∗∗∗ 0.668∗∗∗ 4.459 0.025∗∗∗
Retail trade 0.640∗∗∗ 4.366 0.005∗∗∗ 0.871∗∗∗ 5.995 0.024∗∗∗ 0.833∗∗∗ 5.446 0.035∗∗∗
Finance and real estate 0.141 0.973 0.001 0.253∗ 1.767 0.005 0.433∗∗∗ 2.877 0.015∗∗
Services 0.418∗∗∗ 2.928 0.003∗∗ 0.583∗∗∗ 4.114 0.013∗∗∗ 0.584∗∗∗ 3.909 0.021∗∗∗
Public administration 1.303∗∗∗ 3.411 0.014∗ 1.136∗∗∗ 2.779 0.038∗ 0.383∗∗∗ 0.689 0.014
Constant 1.581∗∗∗ 8.625 1.388∗∗∗ 8.034 1.729∗∗∗ 9.560
No. of firms 192 749 179 860 150 551
Null deviance 79 587 84 320 79 396
Residual deviance 44 328 47 440 43 997
Likelihood ratio test 35 259∗∗∗ 36 880∗∗∗ 35 399∗∗∗
Pseudo R square 0.492 0.493 0.509
Panel B: In-sample prediction tests
One year prior to failure Two years prior to failure Three years prior to failure
No. of firms 192 749 (182 587 active and 10 162 distressed) 179 860 (168 578 active and 11 282 distressed) 150 551 (139 421 active and 11 130 distressed)
Classification accuracy 88.6% (161 681 active firms and 9000 failed 88.5% (149 251 active firms and 9990 failed 89% (124 015 active firms and 9901 failed firms
firms correctly classified using the optimal firms correctly classified using the optimal correctly classified using the optimal cut-off

cut-off point of 0.13) cut-off point of 0.15) point of 0.19)
Area under the ROC 0.946 0.942 0.942
curve
153
(Continued)
This table summarises the results of the full logistic regression models over one- to three-year time horizons. Panel A presents the maximum likelihood estimates of parameters, z-values and marginal
effects. ∗∗∗ , ∗∗ and ∗ denote significance at the 1%, 5% and 10% levels respectively. Panels B and C display in-sample and out-of-sample predictive accuracy in terms of classification of firms and
the area under the ROC curve. The ROC curve plots the true positive rate against the false positive rate at various cut-off scores. Overall, logistic regression predicts financial distress up to three years
The logistic analysis for the second year prior to failure
83.8% (151 207 active firms and 11 076 failed

correctly classifies 88.5% of the firms in the estimation
firms correctly classified using the optimal

193 765 (180 545 active firms and 13 220
sample based on the optimal cut-off point of 0.15 and
achieves an AUC of 0.942. The two-year-ahead out-of-
Three years prior to failure

sample prediction accuracy deteriorates slightly: 84.1%
of the firms were correctly classified using the optimal
cut-off point of 0.06 and an AUC of 0.886. In unreported
0.877
results, the reduced model over a two-year horizon cor-

distressed firms) rectly classifies 67.1% of firms in the test sample and
achieves an out-of-sample AUC of 0.736.
Three years prior to failure, the logistic analysis cor-
rectly classifies 89.0% of the firms in the estimation
sample based on the optimal cut-off point of 0.19 and
achieves an AUC of 0.942. The logistic analysis for the
third year prior to failure achieves worse but still excel-
lent out-of-sample prediction: 83.8% of the firms were
correctly classified based on the optimal cut-off point
219 335 (203 787 active and 15 548 distressed)
of 0.06 and an AUC of 0.877. In unreported results, the

reduced model for the third year prior to failure achieves

much worse out-of-sample predictive accuracy: a cor-
Two years prior to failure
rect classification of 65.5% of firms and an AUC of 0.709.

Overall, the logistic analysis outlined in Panels B and C of
Table 6 predicts financial distress in private companies up
0.886
to three years ahead with excellent discrimination accu-

racy. In contrast, the reduced models without controlling

for number of shareholders and number of subsidiaries
predict financial distress with acceptable discrimination
accuracy only. These results show that adding number
of shareholders and number of subsidiaries significantly
improves the performance of financial distress predic-
tion models both in terms of overall classification success
and AUCs.
232 652 (217 443 active and 15 209 distressed)

Conclusion
One year prior to failure
Financial distress prediction is receiving significant at-

tention because of the increasing incidence of corpo-
rate failures during the GFC. However, the majority of
0.894
this literature has focused on public company failures,

with relatively few studies on private company failures

because of the difficulty in accessing private company
samples. The main aim of this study was to examine the
incremental predictive power of two new non-financial
ahead with reasonably good discrimination.
variables over financial ratios and control variables. In

Panel C: Out-of-sample prediction tests
doing so, the paper contributes to the overall literature

on credit risk modelling for private companies, particu-
larly in relation to alternative predictors.
To achieve the aim, this study developed logistic re-
Area under the ROC curve
gression models to explain the financial distress of pri-

Classification accuracy
Table 6 Continued
vate companies in Italy over one- to three-year time

horizons using a large dataset from the ORBIS database.
Firms were classified into two mutually exclusive cat-
No. of firms
egories: ‘active’ and ‘distressed’. A firm is distressed if

it defaults on payment, or is subject to insolvency pro-
ceedings, bankruptcy or liquidation. To model financial

distress for private companies, several predictors of fi- private companies with a larger number of subsidiaries
nancial distress were selected, including firm age, indus- are also found to be negatively related to private company
try dummies, number of shareholders, number of sub- failures, suggesting that the benefits of a group outweigh
sidiaries, cash to total assets, EBIT to total assets, fixed the costs. The financial and non-financial factors found
assets to total assets, total liabilities to total assets and to influence private company failures can also be used as
working capital to total assets. early warnings so that owner-managers can take correc-
The study presents several key findings. First, it pro- tive or preventive action in a timely manner. The findings
vides insight into how number of shareholders and num- of the study suggest that owner-managers can reduce the
ber of subsidiaries interact with financial distress risk in odds of corporate failures by having other shareholders
private companies, and how these add to the predic- and a group structure.
tive accuracy of models. Over one- to three-year time This study has several limitations. Firstly, the dataset
horizons, number of shareholders and number of sub- used in the study is from 2009–13 and does not include
sidiaries are found to have a negative impact on fi- accounting and non-accounting information of Italian
nancial distress in private companies. The full mod- private companies from 2014 to the present as the ORBIS
els controlling for number of shareholders, number of database is only updated when its data providers submit
subsidiaries and other variables achieve outstanding in- new datasets. Secondly, firms that experience a default
sample discrimination and excellent out-of-sample dis- of payment may not necessarily end up with more severe
crimination. In contrast, the reduced models control- financial distress events such as insolvency proceedings,
ling for other variables achieve acceptable in-sample and liquidation or bankruptcy. A promising avenue for fu-
out-of-sample discrimination only. These findings are ture research is to classify firms into several states instead
indicative of the incremental predictive ability of num- of the conventional active/financial distress dichotomy
ber of shareholders and number of subsidiaries over ac- employed in this study and estimate multinomial logis-
counting ratios and other variables. Second, the study tic regression models accordingly. Thirdly, the models
also documents the relative influence of non-financial estimated in the study control for industry effects by
information. The out-sample prediction tests over a one- incorporating industry dummies. Future research could
year time horizon suggest that number of shareholders explore whether the influence of firm-specific variables
has the strongest overall impact on private company on private company failures varies across industries by
failures, followed by number of subsidiaries, industry estimating industry-specific models.
dummies and firm age.
This study is of interest to a wide range of stakehold-
Notes
ers particularly in the wake of the recent GFC. Firms are
affected by uncertainties arising from changes in exter-
1 Cascino and Gassen (2015) found that public firms adopting
nal economic conditions. ‘Small businesses will be more IFRS become less comparable than private firms adopting local
vulnerable and less able to withstand external economic GAAP in Germany and Italy.
shocks and adverse business conditions’ (McMenamin 2 Prior studies included firm size (natural logarithm of total assets)
1999: 139). By estimating models for the GFC period as a control variable and yet found that the coefficient of firm size
had an unexpected positive sign (e.g., Bhimani et al. 2010). In
and validating them for the post-crisis period, the study
unreported results, the study also found that firm size (natural
yields results which are tailored to Italian private com- logarithm of total assets) had an unexpected positive sign. Ar-
panies in the current economic environment.3 Accord- guably, the natural logarithm of total assets is a poor proxy for
ingly, private company failures modelled as a function firm size: not only are total assets understated due to the influ-
of accounting ratios, non-accounting information and ence of conservatism, but total assets have also been used as the
denominator in several ratios (e.g., cash to total assets, working
in particular number of shareholders and number of
capital to total assets, fixed assets to total assets) used in this study.
subsidiaries have significance to banks and regulators Hence, firm size (natural logarithm of total assets) is not included
that assess credit risk in private companies and overall as a control variable in this study.
financial pressure in these smaller companies. In Italy, 3 The GFC began in 2006 and reached a peak between 2007 and
private companies are typically owned and controlled by 2008; therefore, the 2005–09 period from which the estimation
sample is drawn in this study represents the GFC.
individuals and families who provide protection against
uncertainty through financing, leadership and manage-
ment (Colli 2003). This is consistent with our findings
that private companies owned by a larger number of References
shareholders show lower odds of financial distress. Pri-
vate companies in Italy can be a single company or a Altman, E.I. 1968, ‘Financial Ratios, Discriminant Analysis
corporate group. There are advantages (e.g., financial and the Prediction of Corporate Bankruptcy’, The Journal of
benefits, economies of scale, image and market share) Finance, 23 (4): 589–609.
and disadvantages of a corporate group (e.g., maintain- Altman, E.I., Sabato, G. and Wilson, N. 2010, ‘The Value
ing a complex group) (Pernazza 2010: 239). In this study, of Non-financial Information in Small and Medium-sized

Enterprise Risk Management’, The Journal of Credit Risk, 6 Ciampi, F. 2015, ‘Corporate Governance Characteristics and
(2): 95–127. Default Prediction Modelling for Small Enterprises: An Em-
pirical Analysis of Italian Firms’, Journal of Business Research,
Altman, E.I., Iwanicz-Drozdowska, M., Laitinen, E.K. and Su-
68 (5): 1012–25.
vas, A. 2017, ‘Financial Distress Prediction in an International
Context: A Review and Empirical Analysis of Altman’s Z-Score Ciampi, F. and Gordini, N. 2013, ‘Small Enterprise Default
Model’, Journal of International Financial Management and Ac- Prediction Modelling Through Artificial Neural Networks: An
counting, 28 (2): 131–71. Empirical Analysis of Italian Small Enterprises’, Journal of Small
Business Management, 51 (1): 23–45.
Ang, J.S. 1992, ‘On the Theory of Finance for Privately Held
Firms’, The Journal of Small Business Finance, 1 (3): 185– Colli, A. 2003, The History of Family Business 1850–2000, Cam-
203. bridge University Press, Cambridge, UK.
Balcaen, S. and Ooghe, H. 2006, ‘35 Years of Studies on Business Constable, J.J. and Woodliff, D.R. 1994, ‘Predicting Corpo-
Failure: An Overview of the Classic Statistical Methodologies rate Failure Using Publicly Available Information’, Australian
and Their Related Problems’, The British Accounting Review, 38 Accounting Review, 4 (7): 13–27.
(1): 63–93. Crone, S.F. and Finlay, S. 2012, ‘Instance Sampling in Credit
Beaver, W.H. 1966, ‘Financial Ratios as Predictors of Failure’, Scoring: An Empirical Study of Sample Size and Balancing’,
Empirical Research in Accounting: Selected Studies, Journal of International Journal of Forecasting 28 (1): 224–38.
Accounting Research, 4 (Supplement): 71–111. Edmister, R.O. 1972, ‘An Empirical Test of Financial Ratio
Beaver, W.H. 1968, ‘Alternative Accounting Measures as Pre- Analysis for Small Business Failure Prediction’, The Journal of
dictors of Failure’, The Accounting Review, 43 (1): 113–22. Financial and Quantitative Analysis, 7 (2): 1477–493.
Beaver, W.H., McNichols, M.F. and Rhie, J.W. 2005, ‘Have European Parliament 2017, ‘Small and Medium-Sized
Financial Statements Become Less Informative? Evidence from Enterprises’. Available at: http://www.europarl.europa.eu/
the Ability of Financial Ratios to Predict Bankruptcy’, Review atyourservice/en/displayFtu.html?ftuId=FTU_5.9.2.html,
of Accounting Studies, 10 (1): 93–122. accessed 1 December 2017.
Beck, T., Demirguc-Kunt, A. and Maksimovic, V. 2008, ‘Fi- Falkenstein, E., Boral, A. and Carty, L.V. 2000, ‘RiskCalcTM for
nancing Patterns Around the World: Are Small Firms Differ- Private Companies: Moody’s Default Model’, Working Paper,
ent?’, Journal of Financial Economics, 89 (3): 467–87. Moody’s Investors Service, New York.
Berger, A. and Frame, S.W. 2007, ‘Small Business Credit Scoring Ferguson, A., Clinch, G. and Kean, S. 2011, ‘Predicting the
and Credit Availability’, Journal of Small Business Management, Failure of Developmental Gold Mining Projects’, Australian
45 (1): 5–22. Accounting Review, 21 (1): 44–53.
Berger, A.N. and Udell, G.F. 2006, ‘A More Complete Concep- Filipe, S.F., Grammatikos, T. and Michala, D. 2016, ‘Forecasting
tual Framework About SME Finance’, Journal of Banking and Distress in European SME Portfolios’, Journal of Banking and
Finance, 30 (11): 2945–966. Finance, 64: 112–35.
Bhimani, A., Gulamhussen, M.A. and Lopes, S.D. 2010, ‘Ac- Grunert, J., Norden, L. and Weber, M. 2005, ‘The Role of Non-
counting and Non-accounting Determinants of Default: An financial Factors in Internal Credit Ratings’, Journal of Banking
Analysis of Privately-held Firms’, Journal of Accounting and and Finance, 29 (2): 509–31.
Public Policy, 29 (6): 517–32. Hensher, D.A. and Jones, S. 2007, ‘Forecasting Corporate
Bhimani, A., Gulamhussen, M.A. and Lopes, S.D. 2013, ‘The Bankruptcy: Optimizing the Performance of the Mixed Logit
Role of Financial, Macroeconomic, and Non-financial Infor- Model’, Abacus, 43 (3): 241–64.
mation in Bank Loan Default Timing Prediction’, European Hillegeist, S.A., Keating, E.K., Cram, D.P. and Lundstedt, K.G.
Accounting Review, 22 (4): 739–63. 2004, ‘Assessing the Probability of Bankruptcy’, Review of Ac-
counting Studies, 9 (1): 5–34.
Bhimani, A., Gulamhussen, M.A. and Lopes, S.D. 2014, ‘Owner
Liability and Financial Reporting Information as Predictors of Hosmer, D.W. and Lemeshow, S. 2000, Applied Logistic Regres-
Firm Default in Bank Loans’, Review of Accounting Studies, 19 sion, John Wiley & Sons Inc, New York.
(2): 769–804.
International Monetary Fund 2016, ‘IMF Executive Board
Capriat, M. 2009, ‘The Economic Context’, in J.L. Newell (ed.), Concludes 2016 Article IV Consultation with Italy’. Available
The Italian General Election of 2008: Berlusconi Strikes Back, at: http://www.imf.org/en/News/Articles/2016/07/12/19/59/
Palgrave Macmillan, New York: 43–61. pr16329-IMF-Executive-Board-Concludes-2016-Article-IV-
Consultation-with-Italy, accessed 3 December 2017.
Cascino, S. and Gassen, J. 2015, ‘What Drives the Comparabil-
ity Effect of Mandatory IFRS Adoption?’, Review of Accounting Jones, S. and Hensher, D.A. 2004, ‘Predicting Firm Financial
Studies, 20 (1): 242–82. Distress: A Mixed Logit Model’, The Accounting Review, 79 (4):
1011–38.
Charitou, A., Neophytou, E. and Charalambous, C. 2004, ‘Pre-
dicting Corporate Failure: Empirical Evidence for the UK’, Eu- Jones, S., Johnstone, D. and Wilson, R. 2015, ‘An Empirical
ropean Accounting Review, 13 (3): 465–97. Evaluation of the Performance of Binary Classifiers in the

Prediction of Credit Ratings Changes’, Journal of Banking and Peel, M.J. and Peel, D.A. 1987, ‘Some Further Empirical Evi-
Finance, 56: 72–85. dence on Predicting Private Company Failure’, Accounting and
Business Research, 18 (69): 57–66.
Jones, S., Johnstone, D. and Wilson, R. 2017, ‘Predicting Cor-
porate Bankruptcy: An Evaluation of Alternative Statistical Pernazza, F. 2010, Corporations and Partnerships in Italy,
Frameworks’, Journal of Business Finance & Accounting, 44 (1– Wolters Kluwer, New York.
2): 3–34.
Pindado, J. and Rodrigues, L.F. 2004, ‘Parsimonious Models
Joy, O.M. and Tollefson, J.O. 1975, ‘On the Financial Applica- of Financial Insolvency in Small Companies’, Small Business
tions of Discriminant Analysis’, Journal of Financial and Quan- Economics, 22 (1): 51–66.
titative Analysis, 10 (5): 723–39.
Pompe, P.P.M. and Bilderbeek, J. 2005, ‘The Prediction of
Keasey, K. and Watson, R. 1987, ‘Non-financial Symptoms and Bankruptcy of Small- and Medium-sized Industrial Firms’,
the Prediction of Small Company Failure: A Test of Argenti’s Journal of Business Venturing, 20 (6): 847–68.
Hypothesis’, Journal of Business Finance and Accounting, 14 (3):
Shumway, T. 2001, ‘Forecasting Bankruptcy More Accurately:
335–54.
A Simple Hazard Model’, Journal of Business, 74 (1): 101–24.
Keasey, K. and Watson, R. 1988, ‘The Non-submission
Swets, J.A., Dawes, R.M. and Monahan, J. 2000, ‘Better Deci-
of Accounts and Small Company Financial Failure Pre-
sions Through Science’, Scientific American, 283 (4): 82–7.
diction’, Accounting and Business Research, 19 (73): 47–
54. Thomas, L.B. 2013, The Financial Crisis and Federal Reserve
Policy, Palgrave Macmillan, New York.
Laitinen, T. and Kankaanpaa, M. 1999, ‘Comparative Analysis
of Failure Prediction Methods: The Finnish Case’, European Train, K.E. 2009, Discrete Choice Methods with Simulation,
Accounting Review, 8 (1): 67–92. Cambridge University Press, Cambridge.
McNamara, R.P., Cocks, N.J. and Hamilton, D.F. 1988, ‘Pre- Van Peursem, K. and Chan, Y.C. 2014, ‘Forecasting New
dicting Private Company Failure’, Accounting and Finance, 28 Zealand Corporate Failures 2001–10: Opportunity Lost?’, Aus-
(2): 53–64. tralian Accounting Review, 24 (3): 276–88.
McMenamin, J. 1999, Financial Management: An Introduction, Wei, Y. 2003, Comparative Corporate Governance: A Chinese
Routledge, London. Perspective, Kluwer Law International, New York.
Ohlson, J.A. 1980, ‘Financial Ratios and the Probabilistic Pre- Zmijewski, M.E. 1984, ‘Methodological Issues Related to the
diction of Bankruptcy’, Journal of Accounting Research 18 (1): Estimation of Financial Distress Prediction Models’, Journal of
109–31. Accounting Research, 22 (Supplement): 59–86.


Predicting Private Company Failures in Italy Using Financial and Non-Financial Information

Uploaded by

Copyright:

Available Formats

Predicting Private Company Failures in Italy Using Financial and Non-Financial Information

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Predicting Private Company Failures in Italy Using Financial and Non-Financial Information

Uploaded by

Copyright:

Available Formats

Predicting Private Company Failures in Italy Using Financial

and Non-financial Information

Timothy Wang , The University of Sydney

144 Australian Accounting Review

Table 1 Applications of financial distress prediction models for private firms

ownership of assets, management skills, industry

Predicting Private Company Failures

experience any of the above financial distress events.

Traditional statistical or machine learning

Multiple discriminant analysis and binary

Multi-period logit model and the Cox

proportional hazard model

study uses an unbalanced sample, which better reflects

the actual failure rates. The performance of a logit model

is very robust regardless of whether the model is esti-

was used to exclude observations with no or very lim-

consists of 523 160 firm-years.

an estimated model, a test sample is also developed. A

special cases depending on countries

This produces a test sample of 645 752 firm-years.

(e.g., liquidation in Norway)

The study includes number of subsidiaries and number

dictors, a number of financial ratios used in prior studies

accounting ratios (Balcaen and Ooghe 2006: 84). Fol-

lowing the approach by Filipe et al. (2016), accounting

Altman et al. (2017)

with the dependent variable at a time in univariate analy-

ses, those accounting ratios with an area under the curve

are highly correlated with the correlation coefficient be-

146 Australian Accounting Review

Australian Accounting Review

Table 4 Descriptive statistics for firm-specific variables

150 Australian Accounting Review

Panel A: Estimation results

Model I Model II Model III Model IV Model V

Predicting Private Company Failures

Model I Model II Model III Model IV Model V

correctly classified using the

and 10 764 failed firms

failure by 0.010%. The marginal effects of the industry

dummies are positive. Over a one-year time horizon,

times more likely to fail than those in agriculture, fishing

correctly classified using the

This section explores how effectively the logistic regres-

sion models augmented by number of shareholders and

over one- to three-year time horizons, which is often re-

minimise variance and the OLS approach to goodness-

of-fit is not applicable. To better evaluate the predictive

idates these models on the test sample. The test sample

over one- to three-year time horizons consists of 232

firms based on the cut-off score of 0.5, this study uses

much lower cut-off scores because the models are esti-

mated on the unbalanced sample. This study also uses

first year prior to failure. The model achieves an AUC of

0.946 and thereby provides outstanding in-sample dis-

crimination. The logistic analysis achieves similar one-

the reduced model (i.e., the model without controlling

achieves out-of-sample classification of 69.5% and an