econstor

A Service of

zbw
Leibniz-Informationszentrum
Wirtschaft

Make Your Publications Visible.

Leibniz Information Centre
for Economics

Bernhardsen, Eivind

Working Paper
A Model of Bankruptcy Prediction

Working Paper, No. 2001/10

Provided in Cooperation with:

Norges Bank, Oslo

Suggested Citation: Bernhardsen, Eivind (2001) : A Model of Bankruptcy Prediction, Working

Paper, No. 2001/10, ISBN 82-7553-186-1, Norges Bank, Oslo,
http://hdl.handle.net/11250/2498716

This Version is available at:

http://hdl.handle.net/10419/209799

Standard-Nutzungsbedingungen: Terms of use:

Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Documents in EconStor may be saved and copied for your
Zwecken und zum Privatgebrauch gespeichert und kopiert werden. personal and scholarly purposes.

Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle You are not to copy documents for public or commercial
Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich purposes, to exhibit the documents publicly, to make them
machen, vertreiben oder anderweitig nutzen. publicly available on the internet, or to distribute or otherwise
use the documents in public.
Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen
(insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, If the documents have been made available under an Open
gelten abweichend von diesen Nutzungsbedingungen die in der dort Content Licence (especially Creative Commons Licences), you
genannten Lizenz gewährten Nutzungsrechte. may exercise further usage rights as specified in the indicated
licence.

https://creativecommons.org/licenses/by-nc-nd/4.0/deed.no

www.econstor.eu
ANO 2001/10
Oslo
December 5, 2001

Working Paper
Financial Analysis and Structure Department
Research Department

A Model of Bankruptcy Prediction

by

Eivind Bernhardsen
Working papers from Norges Bank can be
ordered via the Internet:
www.norges-bank.no/english/publications
or from Norges Bank, Subscription service,
P.O.Box. 1179 Sentrum, 0107 Oslo, Norway.
Tel. +47 22 31 63 83, Fax. +47 22 41 31 05

Norges Bank's Working papers present

research projects and reports
(not usually in their final form),
and are intended inter alia to enable
the author to benefit from the comments
of colleagues and other interested parties.

Views and conclusions expressed in Working Papers are

the responsibility of the authors alone.

ISSN 0801-2504
ISBN 82-7553-186-1
 A Model of Bankruptcy Prediction
By
Eivind Bernhardsen
December 5, 2001

Abstract
In this thesis, a model of bankruptcy prediction conditional on financial state-
ments is presented. Apart from giving a discussion on the suggested variables the
issue of functional form is raised. The specification most commonly applied for
the bankruptcy prediction model implies that the rate at which two variables can
substitute another holding predicted risk unchanged will be constant. If the aspect
captured by single financial ratios is considered less a substitute for any other as-
pect as this ratio grows, this restriction may not be appropriate. Specifically, the
structure of constant compensation will make predictions sensitive to non-credible
outliers. A specification of the logit model which allows for flexible rates of compen-
sation is motivated. The model is estimated and the regression results are reported.
Second; by questioning the direct connections between financial ratios and the par-
ticular outcome of bankruptcy, a model structure which determines an upper bound
on probability estimates is explored. By reference to a simple model of misclassifica-
tion, the specification distinguishes between the probability of bankruptcy and the
probability of insolvency. Whereas the predicted probabilities of bankruptcy can be
evaluated empirically, the event of insolvency is not observable. Nevertheless; condi-
tional on the model structure, probabilities can be derived for this event as well. An
evaluation is given on the ability of the model to measure the over-all development
in credit risk for the Norwegian limited liability sector. Individual probabilities of
bankruptcy are multiplied with the firms debt to generate a prediction of expected
loss in absence of recovered values. This measure is then aggregated and fitted
with total loan losses for the Norwegian banking sector over the years 1989-2000.
Finally, the possibility of assessing the eect of macro variables in a short panel of
firms is explored. With reference to an aggregation property of the probit model, a
suggestion is given on how to estimate time-specific eects on aggregate data as a
means to identify macro coecients that can be included in the micro-level model.

Keywords: Bankruptcy, logit analysis, non-linear estimation, aggregation

JEL Classification: G33, C35, C43

i
 Preface:
By this thesis I complete the cand.oecon. degree at the University of Oslo. The thesis
was financed by Norges Bank and was written as a part of a project on the development of
supervisory credit risk models. Participating on this project was inspiring, both professionally
and as a personal experience. I wish to thank the project group; Kai Larsen, Trond Eklund, Terje
Lensberg and Elisabeth Axe. I also wish to thank Bjarne Guldbrandsen and Harald Karlsen for
useful comments. Kjersti Gro Lindquist and Bent Vale from the Research Department provided
me with great support through the process. In particular I wish to thank my supervisor John
Dagsvik.

ii
Contents
1 Introduction 1
1.1 The event of bankruptcy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The decision of continuance . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Restructuring models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 The informational content of the financial statement . . . . . . . . . . . . . 2
1.5 Bankruptcy prediction models . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Methodology 6
2.1 The logit and probit models . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 The random egects probit . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 The method of maximum likelihood . . . . . . . . . . . . . . . . . . . . . . 8

3 A model with flexible rates of compensation 9

3.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.2 Estimation of the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 The data set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3.1 A note on sample selection . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.2 A note on the quality of the data . . . . . . . . . . . . . . . . . . . 15
3.3.3 A note on the panel specification . . . . . . . . . . . . . . . . . . . 15
3.4 The variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.5 Model estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.6 Predictive ability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4 Aggregate predictions 26

iii
5 Financial distress versus bankruptcy 31
5.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

6 A suggestion on further research 35

6.1 Aggregation in the linear regression model . . . . . . . . . . . . . . . . . . 36

6.2 Aggregation in the probit model . . . . . . . . . . . . . . . . . . . . . . . . 36
6.3 Identification of the micro model coecients . . . . . . . . . . . . . . . . . 41

7 Bibliography 42

8 Appendix 44
8.1 Summary statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
8.2 Regression output. Loan losses and risk weighted debt. . . . . . . . . . . . 46

iv
1 Introduction

1.1 The event of bankruptcy

The Norwegian bankruptcy legislation states that a debtor shall begin bankruptcy pro-
ceedings if the debtor is insolvent. The debtor is considered insolvent if he is unable
to fulfill his economic obligations as they mature. He is not considered insolvent if his
property and income are sucient to cover the obligations. The Norwegian penal code
§283a requires a debtor to petition for bankruptcy when the debtor has reason to believe
that the business is run at the expense of the creditors.
Factors which can contribute to the understanding of corporate bankruptcy can be
found both in the fields of Economics and in the theory of Business Management. How-
ever, the many attempts to specify a model of bankruptcy prediction based on causal
specifications of underlying economic determinants has not fully succeeded. The dicul-
ties of merging the theoretical and empirical fields may arise from the diversity of the
phenomenon. Firms are heterogenous and the available information is limited. Further-
more, the event of bankruptcy is twofold as the decision of whether or not to continue
operations is not directly connected to the particular outcome of bankruptcy. In search
of explanatory factors we need not only to identify the factors that influenced on the
insuciency of the firm’s performance, but for the firms that do fail we need to explain
why the particular outcome of bankruptcy was observed, and not a timely liquidation, a
merger, or a restructuring of debt.

1.2 The decision of continuance

If the establishment and abolishment of the firm can be viewed as a reversible investment
decision, or the decision cannot be postponed, at any point in time continuance is optimal
if the present value of operations is in excess of the liquidation value of the firm. This
result is referred to as the standard net present value rule (NPV). If non of the above
conditions hold, NPV need not hold and the decision of continuance is better analyzed in

1
a dynamic framework. The framework of such investment problems is discussed in Dixit
and Pindyck (1994). The option to postpone the investment decision will be valuable,
and should therefore be priced in the alternative cost. Compared with the NPV, at any
point in time a wedge is added to the critical levels of the decision rules. This result is
indeed relevant for the decision of firm continuance: If the entry or exit of markets are
suciently costly and the variance of outcomes suciently high the firm may choose to
operate even at a negative contribution margin.

1.3 Restructuring models

In presence of a positive probability of bankruptcy, the value of a company can be viewed

as a call option which will be valuable to the shareholders only if the market value of
the company is considered greater than the company debt at the date of maturity. If the
option is ”out of money” the creditors will have to bear the loss (i.e. a bankruptcy petition
is filed). The call option need not be exercised, and thus there will be an asymmetry in
the risk faced by shareholders and creditors. This asymmetry may cause the troubled firm
to engage in particular risky projects in egort to recover some value, and so there is a
potential for inecient investment decisions. Models of debt restructuring 1 emphasize the
fact that shareholders, bondholders and debtholders will have digerent priorities on assets
liquidated, digerent ability to control the firm, and digerent exposure to the risk associated
with continuance. By considering digerent assumptions concerning the underlying setting
the restructuring models seeks to analyze what is likely to determine the destiny of a
troubled firm.

1.4 The informational content of the financial statement

The financial statement is a filtered representation of information. Decisions are made

concerning the classification of income and expenses, the timing of income and expenses

1
Restructuring models are analysed in several studies. See for example Myers(1977), Bulow and
Shoven(1978), and Chen, Weston and Altman (1995).

2
as well as the valuation of assets and conventions of depreciation. In many cases the firm
will have incentives to bias the entries; income tax, profit related pay and debt covenant
restrictions are explicitly dependent on the reported figures. The firm may signal profits
to attract investors or to win time in a situation of financial distress.
The use of financial ratios to make qualitative statements about the going concern of
the firm has a long tradition. However, the generality of constructed ratios are controver-
sial. Any textbook of accounting will emphasize the fact that benchmark values are not
directly comparable over digerent industries. Financial ratios must thus be evaluated in
conjunction with additional information related to the nature of the firm and the market
in which it operates: Digerences in trading cycles and degree of capital turnover, mar-
ket competition, volatility of revenues and costs and the industry’s dependency on the
business-cycle are factors of importance.
Moreover; measuring financial ratios is not equivalent with observing ”real character-
istics”, but should rather be considered as ”surrogate measures” of the relevant aspects.
As emphasized by Morris (1989): A unique economic event can result in a variety of
ratio patterns, and a single pattern of ratios can be the result of a variety of underlying
economic conditions (fig 1). The business analyst put on the task of giving a subjective
evaluation of a firm will therefore use the collection ratios interactively. Digerent con-
stellations of the financial entries can give rise to hypothesis of the underlying economic
conditions. Ideally, the analysis is combined with external sources of information so that
an over all profile of the firm can be drawn.
Any statistically derived bankruptcy prediction model implicitly assumes that bench-
marking financial ratios makes sense. The limited success of bankruptcy prediction models
must be viewed in this perspective. Nevertheless, in practice, bankruptcy prediction mod-
els are found useful: The holder of a large portfolio of claims may find it costly to supervise
individual developments and therefore use the credit risk model as a means to make a
first selection of ”follow-ups”. Furthermore the ”objectiveness” of the statistical model
may be appreciated. The ability to discriminate by subjective judgements will potentially
depend grossly on who is making the analysis and his current orientation towards general

3
 Figure 1: The ambiguity of ratio patterns

economic developments. Even if subjective predictions on average are more egective than
those of the statistical model, this source of uncertainty may not be appreciated. If the
degree of accuracy of the statistical model can be accurately measured, the model will be
particularly useful.

1.5 Bankruptcy prediction models

The study of Beaver (1966) is considered the pioneering work on bankruptcy prediction
models. Beaver motivated his model by a framework quite similar to the model of the
gamblers ruin2 . The firm is viewed as a ”reservoir of liquid assets, which is supplied by

2
In the gamblers ruin model one assumes that net assets follows a random walk process with some
fixed probability of a negative cashflow each period. In the case of no access to external capital, the
model is quite simple: For a suciently long sequence of periods there is always some probability for
a clustering of negative cash flows so that the net assets eventually takes on a negative value. For an
application of the gamblers ruin, see Wilcox(1976)

4
inflows and drained by outflows. (...) The solvency of the firm can be defined in terms of
the probability that the reservoir will be exhausted, at which point the firm will be unable
to pay its obligations as they mature”. By this framework beaver state four propositions:

• The larger the reservoir, the smaller the probability of failure.

• The larger the net liquid-asset flow from operations, the smaller the probability of
failure

• The larger the amount of debt held, the greater the probability of failure,

• The larger the fund expenditures for operations, the greater the probability of fail-
ure.

Beaver identified 30 ratios that were expected to capture relevant aspects. By a

univariate discriminant analysis, these ratios were applied on 79 pairs of bankrupt/non-
bankrupt firms. The best discriminators were ”working capital funds flow/total assets”
and ”net income/total assets” which correctly identified 90% and 88% of the cases.
Altman(1968) conducted a similar study applying multivariate discriminant analysis
using the 7 ratios; return on assets, stability of earnings, debt service, cumulative prof-
itability, liquidity, capitalization and size. Applied on 33 pairs of bankrupt/non-bankrupt
firms the model correctly identifies 90% of the cases one year prior to failure.
Ohlson(1980) is the first to apply the logit analysis on the problem of bankruptcy
prediction. By using 105 bankrupt and 2,058 non-bankrupt firms he is also the first to
apply a representative sample. He states that predictive power appears to be less than
reported in previous studies.
Recent years, much attention is given to the choice of methodology. Methods like
recursive partitioning, neural networks and genetic programming are commonly applied
on the bankruptcy prediction problem. Morris (1998) gives a survey on both new and
traditional approaches to bankruptcy prediction.

5
2 Methodology

2.1 The logit and probit models

Assume that the variable yi 5 {0,1} is related to an unobservable index yiW by a linear
function of the explanatory variables xi1 ,xi2 ,...,xik and the random term ui such that:

yiW = q 0 + q 1 xi1 + q 2 xi2 + ,..., + q k xik + ui (1)

yi = 1 if yiW > 0

yi = 0 else

By this structure we have;

P (yi = 1|q 0 xi ) = P (ui > q 0 xi ) (2)

1  F (q 0 xi )

where F () is the cumulative distribution function for u. Most commonly u is assumed

normally or logistically distributed. If u is assumed normally distributed;

0
3qZ xi /j µ 2¶
0 1 t
F (q xi ) = 0.5 exp  dt (3)
(2Z) 2
3"

and the model given by eq(1) is referred to as the probit model. If u is assumed logistically
distributed

1
F (q 0 xi ) = (4)
1 + exp(q 0 xi )

and the model of eq(1) is referred to as the logit. The logistic distribution diger from

6
the normal distribution primarily by being slightly thicker at the tails. The predicted
probabilities will be quite similar unless the sample is large and enriched with observations
at the tails3 .

2.2 The random egects probit

If a panel is constructed one may not want to impose the restriction that u is identically
distributed over observations. In the random egects probit model the error term for
individual i at time t is decomposed such that;

uit = µi + 0it (5)

µi v N(0, j µ )

0it v N(0, j 0 )

cov(µi , 0it ) = cov(µi , x) = cov(0it , x) = cov (0it , 0ir ) = 0

i = 1,2,...,N (6)

t, r = 1,2,..., T , t 6= r (7)

Thus for each individual the element µi is drawn once and added to the constant term.
jµ j
Defining j = j0
such that 4 = j+1
gives the proportion of the total variance contributed
by the panel level variance component then;

Z µ2
i
(n )
"
e3 2j2 Yi

P (yi |xi ) = s F (q 0 xit + µi ) dµi (8)

3" 2Zj t=1
F (q 0 xit + µi ) = x (q 0 xit + µi ) if yit = 1 (9)

F (q 0 xit + µi ) = 1  x (q 0 xit + µi ) otherwise (10)

The hypothesis of 4 = 0 (no individual heterogeneity) can be tested by the likelihood

ratio test. By the restriction 4 = 0 we get the pooled probit model. If 4 > 0 and the

3
An introduction to the logit and probit models as well as the method of maximum likelihood are
given in Maddala (1983)

7
pooled probit is estimated, the estimator of q will still be consistent although inecient.
Estimated standard errors of q will be biased. However, this bias can (ineciently) be
adjusted for by summing within individuals when calculating the standard errors4 .

2.3 The method of maximum likelihood

The model of eq(1) can be estimated by the method of maximum likelihood. By assum-
ing that ui , i = 1,2,...,N is independently distributed the joint probability of observing
{y1 ,y2 ,...,yN } , each element respectively conditional on {q 0 x1 ,q 0 x2 ,...,q 0 xN }, is given by
the likelihood function:

Y
N
yi
L= F (q 0 xi )13yi (1  F (q 0 xi )) (11)
i=1

Maximizing eq(11) with respect to the coecient vector q is equivalent to maximizing

the log of L and is solved by setting

Y ln L
= 0, j = 1,2,...,k (12)
Yq j

Eq(12) will constitute k non-linear equations and must be solved numerically by an

iterative procedure. For the logit and probit models the information matrix given by
Y 2 ln L
YqYq 0
will be positive definite at any stage of the iteration procedure, and the iterations
will converge to a maximum of the likelihood function independently of the initial values
of q.

4
A discussion on the random eects probit model is given in Maddala (1987), and Guilkey and Murphy
(1992).

8
3 A model with flexible rates of compensation

3.1 The model

Applying the specification of (1) for the bankruptcy prediction model, the explanatory
variables xi1 , xi2 , ..., xik would be financial ratios computed from the balance sheet, and
q 1 xi1 + q 2 xi2 +, ..., +q k xik = q 0 xi is a measure of the financial soundness of the firm. If
this measure exceeds a critical value, the firm is assumed to go bankrupt. The critical
value of this variable is assumed to vary among individuals and thus the stochastic term
ui is introduced5 . By any choice of a monotonic distribution function for u this structure
will imply a constant rate of compensation between variables: The change (in units) in
xr needed to ogset a marginal increase in the variable xs such that risk is held unchanged
is independent of the values of xr and xs . By the logit model we have;

1
P (y = 1|q 0 x) = (13)
1 + exp (q 0 x)

The marginal egect of variable xr is given by;

YP
= P (1  P )q r (14)
Yxr

The marginal egect of xr is thus implicitly dependent on xr by the term P (1  P ).

The rate of compensation between xr and xs is given by

YP
Yxr q
=  Yx
YP
s
= s (15)
Yxs |dp=0 Yxr
qr

5
In the following I will supress the subscipt for individuals. The notation xk will refer to the variable
k for any individual i.

9
 In Laitinen and Laitinen (2000) the appropriateness of applying this specification of
the bankruptcy prediction model is questioned. The issue is motivated by the following
numerical example:
Consider first a firm for which the ratios ”cash to total assets” (measuring liquidity)
and ”cash-flow to total assets” (measuring profitability) both are at a 5% level. Assume
furthermore that at these levels the liquidity measure is considered slightly more critical
in such a way that the firm would be considered equally risky at a liquidity level of 3%
if profitability doubled to 10%. Thus;

Y (cash-f low to total assets) 2

= (16)
Y (cash to total assets) |dp=0 5

Consider next a firm with profitability at 5%, and liquidity measured at 50%. If the
same rate of compensation is imposed, a fall in liquidity to 48% would still require a
doubling of the profitability measure if predicted risk is to be held unchanged.
Laitinen et. al. argues that one would not be greatly concerned whether liquidity is
measured at 50% or 48% and thus a constant rate a of compensation appears unreasonable.
The objection could be interpreted in two ways:

• In terms of insolvency risk, the aspect captured by the variable xr is less a substitute
for the aspect captured by variable xs as the variable xr increases.

• The variable xr is considered less likely to be a relevant measure of the target aspect
as the variable is measured at more ”extreme” values.

By the first suggestion we would like the marginal egect of the variable xr to decline
with the level of xr at any given P . If the second suggestion is considered relevant we
would generally like the marginal egect to decline as the variable deviates from some
critical value. Indeed this a relevant issue. Financial ratios are artificially generated by
the division of entries like ”total assets” or ”revenue from operations” as a means to adjust
for size or the level of activity. These are rough measures, and will be highly irrelevant for

10
some firms. If the aspect is irrelevantly measured the variable is likely to take odd values.
Most studies conducted on large samples typically apply truncations of the variables at
some quantile. However, the significance of the variable will potentially depend on the
choice of truncation and thus make it dicult to determine which set of ratios to include
in the model. Furthermore, the optimal truncation of one variable may depend on what
truncations are applied on the other variables.
Laitinen et. al. does not impose a specific structure, but rather leaves the question
of functional form open. A Taylor expansion of the underlying functional relationship at
the mean values of the variables is used to motivate the inclusion of cross products and
squares of variables in the logit model.
An alternative to this approach would be to impose a specific functional form with the
desired properties. If interaction egects between variables are ignored we could consider
the model;

y W = q 0 + q 1 T1 (x1 ) + q 2 T2 (x2 )+, ..., +q k Tk (xk ) + u (17)

y = 1 if y W > 0 (18)

y = 0 else (19)

,where the function Tr (xr ) is possibly non-linear in the explanatory variable xr . One sug-
gestion for Tr (xr ) would be to apply the Box-Cox transformation;

(xr )br  1
Tr (xr ) = (20)
br

For b equal to one the transformation is linear. For br < 1 , the transformation will
be concave and thus at any level of P the marginal egect of variable xr will decrease with
the level of xr . If br > 1 the transformation is convex. By applying this transformation,
if both br and bs is found not to be significantly digerent from one, this would suggest
that the rate of compensation between xr and xs is constant. However, the Box-Cox
transformation is not defined for negative values of x. This will be problematic for the

11
bankruptcy prediction model as the cash-flow measure potentially will take on negative
values. Furthermore, assuming a concave relationship, at any given level of P the marginal
egect of xr will be greatest at xr = 0. More generally we might like the marginal egect
to decline as xr deviates from some constant kr . If we impose the cumulative logistic
function for Tr (xr ) , by estimating the scale and location parameters specific for each
variable we will have this structure:

1
Tr (xr ) = , i = 1, 2, ..., k
3( xrB3kr )
(21)
1+e r

kr = location parameter f or variable xr

Br = scale parameter f or variable xr

This transformation does not include the linear function as a linear case but on a given
interval for xr , the function Tr (xr ) will be approximately linear if the scale parameter B r
is suciently large. To illustrate this, a plot of Tr (xr ) for xr 5 [100, 100] , kr = 0, and
B r = 10 , B r = 10 and Br = 100 respectively, is given in fig2.

1 δ=1
δ = 10
0.8
δ = 100
0.6

0.4

0.2

-100 -50 0 50 100

Fig 2: T(x) with various scaling

For a suciently large Br , the transformation can roughly be viewed as a re-scaling of

12
the q-coecient6 ;

Tr (xr )  ar + br xr (22)
YP
,  P (1  P )br q r
Yxr

For suciently large scale parameters B r and B s the rate of compensation between the
variable xr and variable xs will thus be constant.
In general, the rate of compensation between the variable xr and variable xs will not
be constant, as the marginal egect of the variable xr will depend explicitly on the value
of xr :

YP q
= P (1  P )Tr (xr )(1  Tr (xr )) r (23)
Yxr Br

When xr = kr , Tr (xr ) will equal 0.5 and thus Tr (xr )(1  Tr (xr )) will be at its maximum.
Independently of B r , the position parameter kr thus determines which value of xr maxi-
YP
mize Yxr
for a given probability P . As the variable xr deviates from kr , Tr (xr )(1  Tr (xr ))
will approach zero. The rate of compensation between xr and xs is given by:

Yxr q Ts (xs )(1  Ts (xs )) B r q B

= s  grs (xr , xs ) s r (24)
Yxs |dp=0 q r Tr (xr )(1  Tr (xr )) B s q r Bs

q s Br
For given a given value of q r Bs
, the change in xr needed to compensate a rise in xs
will be larger the more xr deviates from kr , and smaller the more xs deviates from ks .

6
In this case, only the product b and not the  coecient will be identifiable, . Furthermore, it will
not be possible to separate a from the constant term of the model.
r
1 1 r
ar = r , bi = µ r
¶2 e
r
1+e r 1+e r

13
3.2 Estimation of the model

The model can be estimated by the method of maximum likelihood. Conditional on an

objective set of initial values of (k, B) one can switch between estimating q conditional
on (k, B) and (k, B) conditional on q. By setting suciently large initial values of B the
procedure will have the approximately linear transformations as a starting point.

3.3 The data set

The data used in this study was constructed by the SEBRA-database at Norges Bank. The
database contains the annual financial statements of all limited liability firms registered at
the Norwegian register for business enterprises over the years 1988-1999. The bankruptcy
data is computed by Dun and Bradstreet, and is more or less complete form 1990-1999.
In the preliminary examination of the data it was found that for most bankrupt firms,
there existed a substantial lag between the date of the last registered financial statement
and the date of bankruptcy: If the last registered statement were recorded in year t ,only
25% of the bankrupt firms are declared bankrupt in year t+1, 55% in year t+2, and 20%
in year t+3. Because of this feature of the data, it was decided to use only the years
1990-1996 for estimation. Furthermore it was considered most appropriate to define the
endogenous variable by the event ”the firm was registered bankrupt within 3 years and
this year constitutes the last registered financial statement”. By this approach a pooled
panel structure could be estimated without multiple counting of the responsive event.
Examining this sub-sample some observations were excluded due to a missing bankruptcy
variable. Furthermore, firms for which the book value of total assets did not exceed
250.000 NOK were excluded: For these firms the entries of the financial statements were
frequently considered dicult to interpret and thus suspected to be plagued with errors
of registration. The estimation sample was constructed by the remaining sample, now
containing 398.689 observations including 8.436 bankruptcies.

14
3.3.1 A note on sample selection

The number of registered bankruptcies in the SEBRA database is far less than the number
found in the ocial statistics on the Norwegian limited liability sector7 . The SEBRA
database only includes firms for which the financial statement some year was approved by
the Register for Business Enterprises8 . If the financial statement of a newly established
firm is more likely to be disapproved when the bankruptcy risk is high, this will generate
a sample selection problem.

3.3.2 A note on the quality of the data

Quite frequently, firms were found to be temporarily absent from the data-set, and the
number of firms absent showed significant variation over the estimation period. Further-
more; the year 1994 surprisingly contained a very small number of new establishments
(about a tenth of the sample average).
The change in the proportion of bankruptcies recorded in the SEBRA database did
not show strict correspondence with ocial statistics, sometimes not even in signs. The
bankruptcy data was gathered from a digerent source than the financial statements, and
the quality of this variable is suspected vary over the estimation period.
The financial statements recorded in the SEBRA-database are adjusted prior to the
year of 1992, as an attempt to incorporate the egects of the 1992 Norwegian tax-reform.
The risk of adverse egects due to time-specific sample features was believed to be sub-
stantial. A pooling of the data as a means to smooth the sample was therefore preferred.

3.3.3 A note on the panel specification

Tentatively, the random egects probit was estimated. By the likelihood ratio test the
restriction 4 = 0 could however not be rejected. This was not taken as evidence of

7
The number of bankruptcies recorded in the SEBRA database was compared to the ocial numbers
of Statistics Norway. On average the number reccorded in the database is lower by 30%.
8
Foretaksregisteret Brønnøysund

15
absence of individual heterogeneity, but rather as a result of the consistency property of
the pooled probit specification.
Considerable egort was made to explore whether a dynamic specification could show
useful. However; even for the sample where only firms that were present at t  1 were
included, lagged variables and lagged probability predictions (quite surprisingly) showed
little significance. Some success was found for dummy variables that captured events like
”revenues did drop more than 20% and short term debt did rise”. However, the success
was limited and it was not considered practical to include these variables.

3.4 The variables

• Liquidity:

Cash and deposits - Value of short term debt

lik = Revenue from operations
Outstanding payments of public dues
ube = Total assets
Trade creditors
lev = Total assets

• Profitability:
Result before extra ordinary items + Ordinary write ogs + Depreciation - Taxes
tkr = Total assets

• Solidity:
Book value of equity
eka = Total assets

taptek =”Current book value of equity is less than the value of equity injected” (dummy)

div =” Dividends paid current year” (dummy)

• Age:

aX =”Number of years since incorporation ”, x = 1,...,8 (dummies)

• Size:

size = (ln(Total assets)  8.000)2

16
 • Industry characteristics9 :

meanek = Mean value of the variable eka

meanlev = Mean value of the variable lev

sdtkr = Variance of the variable tkr

The list of explanatory variables applied must be viewed as a suggestion. The vari-
ables named ”lik”, ”tkr” and ”eka” are traditionally used for the analysis of credit risk
at Norges Bank. In this thesis these variables are used as core measures of liquidity,
profitability and solidity. The remaining variables were found by trial and error10 . By the
number of observations in the estimation sample, one would expect that some generality
can be assumed for these variables. The comment on these variables should however be
viewed as suggestive.

• Liquidity:

The amount of cash the firm needs to service its going expenditures will depend
fundamentally on the nature of activities, and one should be reluctant to consider
benchmark values for liquidity ratios. However; commonly firms are drained in terms
of liquid assets immediately prior to bankruptcy, and it may borrow heavily to man-
age its short term obligations. Commonly, ratios like ”short term debt to revenue
from operations” and ”cash to total assets” are found useful in bankruptcy pre-
diction models. In the credit risk model of Norges Bank the aspect of liquidity
is sought captured by the variable ”cash minus short term debt to revenue from
operations”. Applying this variable is analogous with the inclusion of both ”short
term debt to revenue from operations” and ”cash to revenue from operations” if a
coecient restriction is imposed11 .

9
The 5-digit industry code of Statistics Norway was used. The degree of crudnes of this clasification
was determined as to include at least 1000 firms.
10
Summary statistics on all variables are reproduced in the appendix.
11
Empirical support was not found for this restriction. The restiction was however not found to
significantly aect the predictive power of the model, and was applied mainly for practical reasons.

17
• Profitability:

The profitability of the firm should be considered the driving factor for both the
liquidity and solidity aspect. In the long run, the firm must generate a sucient
margin on its operations to be able to service its debt. Sustained negative profits
will quickly drain the solidity of the firm, and if the firm is to expand it may need to
retain earnings in excess of existing requirements. In the short run, negative profits
will quickly drain the liquidity of the firm. Furthermore; the profitability of the
firm is likely to influence the ability of obtaining external finance. The aspect of
profitability is sought captured by a straight forward measure of return on capital
employed.

• Solidity:

If markets are not perfect, the capital structure will be of importance for the con-
tractual relationship between shareholders and debtholders. The greater the share
of shareholders equity, the lower the financial risk, and the firm is more likely to ob-
tain external finance. The book value of equity is a residual measure in the balance
sheet, and thus directly related to the valuation of the firms assets. Furthermore;
the equity share of total assets will give information on the historic performance,
and serve as a buger on future negative profits.

• Outstanding public dues to total assets:

Often bankruptcy proceedings will be initiated by a bankruptcy petition submitted

by the revenue authorities. The authorities have definitive procedures for treating
default payments on taxes and dues, and will generally not negotiate with an insol-
vent debtor. It is reasonable to expect that the debtor will give priority to these
obligations. Thus, if public dues are used as a liquidity buger the firm is likely to
be in severe distress.

18
• Trade creditors to total assets:

The ”natural” level of trade creditors to total assets will vary extensively among
industries, and thus any egort to benchmark this variable is controversial. However,
by including both an industry variable (see below) and an individual variable that
seek to capture this aspect one can hope to establish whether trade creditors is used
as a buger on liquidity.

• ” Book value of equity is less than injected equity”

This variable may indicate to what extent a given level of equity to total assets is
the result of accumulated earnings.

• ”Dividends are paid current year”

Dividends may be used to signal profitability or, if the firm is troubled, as a means
to withdraw assets from the creditors. For these reasons one can easily question the
usefulness of including this variable in a model of bankruptcy prediction. However,
the Norwegian legislation on limited liability companies states that dividends are
not to be paid if there is reason to believe that the firm is in risk of immediate
insolvency. If the legislation is obeyed, the variable should serve as a signal of
solidity.

• Industry mean of equity to total assets:

If the variable ”equity to total assets” is most properly measured by its deviance
from the industry mean this variable will show significant. If only deviance from
industry mean matters, the coecient on this variable should have the opposite
sign of the coecient on equity to total capital. However, the variable was found
negatively correlated with bankruptcy. Accordingly this variable contributes with
some additional information concerning the risk related to the industry. One should
be careful to give definitive interpretations of this result. However, since solidity is
partly a result of retained earnings, one could suspect that industries characterized
with high leverage are subject to more competition than industries with low leverage.

19
 In an industry with a high degree of competition we would expect that both entry
and exit rates are likely to be higher.

• Industry mean of trade creditors to total assets.

If trade creditors to total assets is most properly measured by its deviance from
industry mean, the variable will show statistically significant. However; the vari-
able was found positively correlated with bankruptcy and does thus appear to give
additional information concerning the risk of the industry. The result may capture
the fact that restaurants and retail business are associated with both a high level of
trade creditors and high bankruptcy rates. A dummy variable for restaurants could
not compensate for the exclusion of this variable.

• Industry variance of return on capital employed.

Economic risk will be reflected in the variability of a company’s earnings over time.
If the industry is associated with a high level of variability of earnings and thus a
high level of risk, we would expect both a higher rate of prosperous firms as well as
a higher rate of bankruptcies.

• Number of years since establishment (dummy variables for each of the first 8 years).

Uncertainty concerning the true costs of production as well as factors concerning

the competitive setting makes of establishment of a business risky. Furthermore; the
firm may need time to develop a functional organizational structure and sucient
management skills.

• Size of the firm:

The size of the firm is commonly identified as a significant factor in bankruptcy pre-
diction models. Commonly the logarithm of total assets is employed. In this study
this variable was not found significant. However the square deviance from 2 mill
NOK did enter significantly. This result may indicate that if the firm is suciently
small, (administrative) bankruptcy costs will exceed the expected liquidation value
of the firm, and thus the creditor may not want to initiate bankruptcy proceedings.

20
3.5 Model estimates

When estimating the model initial values were respectively set such that kr = 0, Br = 1,
r = 1, 2, ..., k (Model A) and kr = 0, B r = 100, r = 1, 2, ..., k (Model B). As the
variables were measured in percent, the interval [100, 100] captured at least 98 percent
of the observations for any variable. With reference to fig 2; by setting B r = 100 the
variables will enter approximately linearly in eq(17) at the start of the iteration process.
Table 1. Model(A) estimates:
k 1
variable q s.e B
s.e. B
s.e
eka -1.4459 0.0604 0.4464 0.0977 0.0782 0.0049
tkr -1.0948 0.0386 0.1216 0.1274 0.2096 0.0190
lik -1.4925 0.0421 -2.9618 0.1977 0.1529 0.0087
lev 0.4968 0.0486 1.5224 0.4142 0.2895 0.0660
ube 6.8069 0.2019 -1.1474 0.0243 0.0362 0.0017
a1 0.8380 0.0438 ............ ........... ........... ............
a2 0.9707 0.0382 ............ ............ ............ .............
a3 0.8310 0.0398 ............ ............ ............ ............
a4 0.6729 0.0429 ............ ............ ............ ............
a5 0.5282 0.0468 ............ ............ ............ ............
a6 0.3189 0.0528 ............ ............ ............ ............
a7 0.2689 0.0575 ............ ............ ............ ............
a8 0.2076 0.0638 ............ ............ ............ ............
div -1.0639 0.0742 ............ ............ ............ ............
taptek 0.5386 0.0419 ............ ............ ............ ............
size -0.0543 0.0064 ............ ............ ............ ............
meanlev 1.0404 0.1692 ............ ............ ............ ............
meanek -3.9690 0.2273 ............ ............ ............ ............
sdtkr 1.8229 0.3319 ............ ............ ............ ............ LR chi(19) = 21909.7

const -7.0131 0.2786 ............ ............ ............ ............ log-L= -29917.726

21
 Table 2. Model(B) estimates:
k 1
variable q s.e B
s.e. B
s.e
eka -1.4006 0.0589 0.3594 0.0987 0.0791 0.0051
tkr -1.0594 0.0381 -0.0188 0.1301 0.2115 0.0195
lik -1.3629 0.0391 -3.5081 0.2447 0.1704 0.0107
lev 1.4287 0.1182 -0.5975 0.1169 0.1670 0.0286
ube 2.7226 0.0812 0.7715 0.0418 0.0389 0.0016
a1 0.8297 0.0438 ............ ........... ........... ............
a2 0.9735 0.0382 ............ ............ ............ .............
a3 0.8310 0.0398 ............ ............ ............ ............
a4 0.6753 0.0430 ............ ............ ............ ............
a5 0.5312 0.0469 ............ ............ ............ ............
a6 0.3204 0.0528 ............ ............ ............ ............
a7 0.2709 0.0575 ............ ............ ............ ............
a8 0.2083 0.0639 ............ ............ ............ ............
div -1.0826 0.0742 ............ ............ ............ ............
taptek 0.5496 0.0419 ............ ............ ............ ............
size -0.0573 0.0064 ............ ............ ............ ............
meanlev 1.2807 0.1690 ............ ............ ............ ............
meanek -3.9623 0.2280 ............ ............ ............ ............
sdtkr 1.8229 0.3323 ............ ............ ............ ............ LR chi(19) = 21846.80

const -3.6069 0.2458 ............ ............ ............ ............ log-L = -29932.503

Analytically I have not explored whether the maximum likelihood problem of deter-
mining scale and position parameters conditional on the q estimates has a global maximum
independent of the initial values. The estimates diger slightly, but the model estimates
can not be considered crucially dependent on the choice of initial values.
The variables ”equity to total assets” and ”public dues to total assets” enter quite
linearly in eq(17) with the scale parameters 12. 642 and 41. 667 respectively. In contrast,
the profitability variable enters with B = 4. 728 1. In Fig 3, marginal egects are simulated
for Model (A). The simulation is conducted by setting all variables (except for the one

22
plotted) at their mean values. The simulated firm has been operating for 3 years, did not
experience loss in equity and did not pay dividends the relevant year. The profitability
variable is egective in a quite narrow interval only. In fact; the model was found to
perform surprisingly well when this variable was replaced with a dummy variable for
negative profits.

Marginal effects

0.0002 ube
lev
-100 -50 50 100
0
Ratio in percent
-0.0002
eka
-0.0004
lik
-0.0006
tkr
-0.0008

-0.001

Figure 3. Simulated marginal egects on the probability of bankruptcy.

The rates of compensation between various variables are explored in fig 4. The figure
to the left gives the rate at which a fall in solidity compensates a marginal rise in liquidity.
The figure to the right gives the rate at which a fall in liquidity compensates a marginal rise
in profitability. In the saddle point the marginal egects of both variables are maximized
at any given probability p.

23
 4
4

rate rate
2
2

0
-40 -30 -10
-20 0 -20
liquidity-10 solidity profitability0 liquidity
0 10 20 10
0

Fig 4: Simulated rates of compensation.

3.6 Predictive ability

If the model is to be used to make binary predictions, a cut-og point for the predicted
probability must be determined. The optimal cut-og point will depend on the relative
cost of type one and type two errors. Figure 5 gives the menu of trade-ogs between
correct classifications of the bankrupt cases (sensitivity) and incorrect classifications of
the non-bankrupt cases (1-specificity). The area under the curve above the 45 degree line
is a common measure of discriminatory power. For the model with no explanatory power
the area under the curve will equal 0.5. If predictions are perfect the measure will equal
1.

24
 Figur 5. Discriminatory power

Applied on the estimation sample, the model can correctly identify 83% of both the
bankrupt and non bankrupt cases. When the model is estimated over the years 1990-
1993 it correctly classifies 82% of both categories in 1996. Only 3% of the observations
constitutes bankruptcies. Incorrect classification of 17% of the non-bankrupt cases there-
fore gives a great number of false predictions. If the cut-og point is determined as to
equal the 90% percentile, such that 10% of the sample is classified as bankrupt, one will
correctly classify only 63% of the bankrupt cases. However, because the metric subject
to prediction is defined by the event ”this is the last registered financial statement, and
bankruptcy is recorded within 3 years” a firm which is predicted bankrupt in year t is
noted as an erroneous prediction if year t does not constitute the last financial statement,
even if bankruptcy is recorded within year t + 3.
Figure 6 gives a picture of model stability. The model is estimated over the years
1990-1993 and 1990-1996 respectively and the predicted probabilities are plotted. The
predicted probabilities in general and the ranking of the firms in particular is not notably
agected by this extension of the estimation sample.

25
 Figure 6. Model stability.

4 Aggregate predictions

In table 3, groups are defined by various intervals of predicted probabilities and mean pre-
dicted risk is compared with the observed fractions by year12 . The first order condition for
maximizing the log likelihood for the logit model assures that mean predicted probability
coincides with proportion of responses in the estimation sample. However, this does not
need to hold for every quantile of the predicted probabilities. Table 3 suggests that pre-
dicted probabilities fit well with the observed frequencies, not only over the distribution
of predicted probabilities, but also over the digerent years. Despite the digerences in the
business cycle, there appears to be some degree of stability in the correspondence between
mean predicted risk and the observed frequencies over the years 1990-1996.

12
For the group of firms with a predicted probability of bankruptcy in excess of 20% in 1990, the
frequency of bankruptcies was 31 % . For the same group, mean predicted probability were 29%. For the
same year; among the firms with a predicted probability less than 1 percent, the frequency of bankruptcies
was 0.4 %. Mean predicted probability for this group equaled 0.3 %.

26
Table 3. Mean predicted risk and observed fractions

Cut-o point Fraction 1990 Mean predicted 1990 Fraction 1991 Mean predicted 1991

p>.2 .315827 .2939747 .309848 .2976207

.2>p>.1 .186139 .1409671 .170139 .1400238

.1>p>.05 .107498 .0710920 .100804 .0708527

.05>p>.02 .049738 .0322007 .048263 .0322326

.02>p>.01 .021866 .0143543 .019082 .0142045

p<.01 .004291 .0032402 .004241 .0032123

Cut-o point Fraction 1992 Mean predicted 1992 Fraction 1993 Mean predicted 1993

p>.2 .234043 .2933678 .299776 .2920722

.2>p>.1 .119509 .1389947 .134093 .1390491

.1>p>.05 .065691 .0703857 .071406 .0700618

.05>p>.02 .026517 .0321568 .025918 .0318994

.02>p>.01 .010394 .0142870 .011211 .0142867

p<.01 .002148 .0028966 .002405 .0027531

Cut-o point Fraction 1994 Mean predicted 1994 Fraction 1995 Mean predicted 1995

p>.2 .271605 .2947980 .188335 .2896208

.2>p>.1 .123089 .1378097 .107410 .1387632

.1>p>.05 .055202 .0698960 .057420 .0703471

.05>p>.02 .024048 .0318248 .027301 .0317281

.02>p>.01 .010627 .0142041 .009455 .0142126

p<.01 .001627 .0026003 .001434 .0026231

Cut-o point Fraction 1996 Mean predicted 1996

p>.2 .231141 .2955294

.2>p>.1 .123243 .1390570

.1>p>.05 .065801 .0707439

.05>p>.02 .029142 .0317599

.02>p>.01 .014248 .0141861

p<.01 .002258 .0025373

27
 This result is particularly useful for the egort of predicting expected loss on a port-
folio of firms. Figure 9 gives a plot of risk weighted debt 1988-1999. This measure is
constructed by multiplying individual probability predictions with the firms debt, and
can be interpreted as expected loss related to the event of bankruptcy on the SEBRA
portfolio in absence of collateral values to be recovered. Figure 10 gives the reported loan
losses of the entire Norwegian banking sector for the same period13 . Figure 11 gives the
fitted regression of loan losses on risk weighted debt with a one year lag. Risk weighted
debt is lagged for two reasons. First; the analysis on micro data identified a substantial lag
between the last reported statement and the date of bankruptcy for most bankrupt firms.
Second; we would expect banks to analyze the statements of year t-1 when determining
provisions on loan losses at time t. As the financial statements for year t is available as late
as in the middle of year t+1, we need this lag to be able to make predictive statements.

Risk weighted debt, millions NOK. Constant 2000 prices

6000

5000

4000
Risk weighted debt,
3000 millions NOK.
Constant 2000 prices
2000

1000

0
88

90

92

94

96

98
19

19

19

19

19

19

Figure 9

13
Loan losses of the year and loanloss provisions minus write-backs of previous years loanloss provisions.

28
 Loan losses. Millons NOK, constant 2000 prices

30000

25000

20000

15000 Loan losses. Millons

NOK, constant 2000
10000 prices

5000

-5000
88

90

92

94

96

98

00
19

19

19

19

19

19

20

Figure 10

30000

25000

20000

15000

10000

5000

-5000
88
89
90
91
92
93
94
95
96
97
98
99
00
19
19
19
19
19
19
19
19
19
19
19
19
20

Fig 11. loan losses and fitted loan losses

The R-squared of the regression with a constant term included is 81%. In particular,
the regression does fit well with the massive loan losses in 1991. Although one should be
careful to draw conclusions on the basis of such a short time-series, this finding is indeed
encouraging. Several studies have explored the dependence of aggregate loan losses on

29
various macro variables14 . However, to the knowledge of the author, not many attempts
has been made to explain bank sector loan losses by aggregating micro predictions.
Because there is a tendency of under-prediction at the early years and over-prediction
at late years one would expect that inclusion of a macro-variable will show useful in this
model. In particular we would like to include a variable that captures the variation in
collateral values over the business cycle. In fig 12 loan losses is fitted with lagged risk
weighted debt and the change in the real price of housing (the econ index). The macro
variable does improve the fit of the model15 .

30000

25000

20000

15000

10000

5000

0
1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000
-5000

Figure 12. Loan losses fitted with risk weighted debt and change in the econ real price
of housing index.

14
See Pesola (2000).
15
The regression output is reproduced in the appendix.

30
5 Financial distress versus bankruptcy

5.1 The model

The usual specification of the bankruptcy prediction model implicitly assumes that the
event of bankruptcy is directly connected to the quality of the financial statements: If
the figures are suciently bad, we expect the firm to go bankrupt. Yet, as emphasized
by the restructuring models, once in a situation of financial distress bankruptcy is only
one of several possible outcomes.
Assume that q 0 x measures the financial soundness of the firm, and furthermore that
there exist a critical level of this measure such that if this level is exceeded the firm has
an insolvency problem. Once in the state of insolvency, assume the claimants on the firm
will initiate negotiations, and furthermore that the outcome of these negotiations cannot
be predicted by financial statements. Formally: let the event of insolvency be represented
by a the binary variable y o , and assume that the outcome of bankruptcy will occur with a
fixed probability conditional on this variable. The critical level of q 0 x is assumed to diger
among the firms due to individual characteristics that is not captured by the financial
ratios, and thus the error term u is introduced:

yW = q 0 x + u (25)

y o = 1 if y W > 0

y o = 0 else

In this model not only y W is latent, but also the variable y o . What is observable is the
outcome of bankruptcy represented by the binary variable y, such that:

P (y = 1|y o = 1) = (1  q) (26)

P (y = 0|y o = 0) = (1  r)

31
 Thus:

if y W > 0 then y = 1 with probability (1  q)

if y W  0 then y = 1 with probability r

Solving the model for the probability of observing bankruptcy we have:

P (y = 1|x) = P (y = 1|y o = 1) P (y o = 1) + P (y = 1|y o = 0) P (y o = 0) (27)

= r + (1  q  r) P (u > q 0 x)

Assuming u is logistically distributed we have;

(1  q  r)
P (y = 1|x) = r + (28)
1 + exp(q 0 x)

By this model, the probability of observing bankruptcy will be constrained to the

[r, (1  q)] interval. Conditional on the model structure, the identification of q allows
consistent probabilities of the (unobservable) event of financial distress to be calculated.
Independently of how one would interpret the motivated setting, the proposed functional
form may be desirable: The transformation of variables suggested in section 3.1 implied
that the marginal egect of a single variable should decline as the variable deviates from
some critical value at any given level of the probability of bankruptcy. The functional
form suggested in this section will have implications as to how the marginal egect of any
variable is related to the over-all evaluation of the firm: By the structure given in section
3.1, the marginal egect of any variable xr was given by;

YP q
= P (1  P )Tr (xr )(1  Tr (xr )) r
Yxr Br

The marginal egect is dependent on the over-all evaluation of the firm by the term

32
P (1  P ) ,which is maximized at P = 0.5. By the model of eq(28) , if we impose the
restriction of r = 0, the marginal egect of variable xr is

YP P q
= P (1  )Tr (xr )(1  Tr (xr )) r
Yxi 1q Br

At any given level of xr the marginal egect will be greatest when P W = 0.5(1  q). If
q > 0 (still we impose r + q < 1), this will imply that marginal egects will be maximized
at lower level of P , and thus the over  all evaluation of the firm contributes to marginal
egects more conservatively as probability estimates grow large.

5.2 Estimation

16
The structure is analogous to a basic model of misclassification : The log likelihood is
given by:

; ³ ´ <
n ?
X (13q3r) @
yi ln r + 1+exp(3q 0
x)
)
` ((q, r, q) = n31 ³ ³ ´´ (29)
= + (1  y ) ln 1  r + (13q3r)0 >
i=1 i 1+exp(3q x)

For the identification of the vector (q, r, q), the condition q + r < 1 must be imposed, as
the estimators based on the maximum likelihood procedure will not be able to distinguish
between the parameter values (r, q, q), and (1  r, 1  q, q)17 . The model is estimated
on transformed variables, conditioned on scale and position parameters identified in the
previous section Model(A).

16
J.A. Hausman et al (1998)
17
It was found dicult formally to impose this restriction on the standard Stata ML-evaluators. Tech-
¡s ¢2
nically the restriction was imposed by replacing 1  r  q with 1  r  q computing the log-likelihood
equation. In this way the program was kept from evaluating the log likelihood at r + q > 1

33
Estimation results. Model C.

variable coeficient s.e.

eka 1.6161 0.0696
lik 1.6337 0.0498
tkr 1.3022 0.0490
lev 0.5111 0.0534
ube 8.2014 0.2823
a1 0.9349 0.0513
a2 1.1107 0.0472
a3 0.9375 0.0475
a4 0.7416 0.0496
a5 0.5723 0.0531
a6 0.3361 0.0588
a7 0.2814 0.0636
a8 0.2058 0.0702
div 0.9756 0.0752
size 0.0511 0.0069
taptek 0.5012 0.0438
meanlev 0.9192 0.1921
meaneka 4.8429 0.2902
sdtkr 2.3266 0.3804
const 6.7421 0.3244
q 0.5110 0.0286 Wald chi(19) = 4492.89

r 0.0000783 0.000379 log-L = -29847.179

34
 The probability of bankruptcy given insolvency is estimated to equal 49%, which
appears to be reasonable. The probability of bankruptcy given solvency is not significantly
digerent from zero at a 1% level. By this model the marginal egect of any variable x is
maximized at P = 0.245 for any given level of the same variable x. Figure 13 gives a plot
of model A against model C predictions. By the likelihood ratio test, the joint restriction
r = q = 0 (model A against model C) is rejected.

Fig 13 A plot of model A and model C predictions.

6 A suggestion on further research

For most panels of accounts data the number of years will be small and estimating the
egects of macro variables will be dicult. Sometimes a longer time-series will be available
on aggregated data. In what follows I will explore whether a model estimated on aggre-
gated data can be utilized to identify macro-coecients for inclusion in the micro-level
model.

35
6.1 Aggregation in the linear regression model

To set focus, I will quickly review aggregation in the case of the linear regression model:
Assume the model:

yit = k + xit q + zt  + uit , i = 1, 2, ..., N , t = 1 (30)

With nothing but cross-sectional data available we cannot separate  from the constant
term. Assume however that a time-series is available on both the time specific variable
and the mean value of x. By aggregating over individuals in eq( 30 ), we have

yt = k + xt q + zt  + ut , t = 1, 2, ..., T (31)

the ”between period” estimate of q. Although we are subject to cross-sectional data on

the individual characteristics, by including the estimate of  in eq(30) and adjusting the
constant term we can now obtain individual predictions conditioned on future development
in the time specific variable, also for the case in which x and z are correlated.

6.2 Aggregation in the probit model

Consider the usual pooled probit specification;

yitW = xit q + zt  + uit i = 1, 2, ..., N ,t=1 (32)

yit = 1 if yitW > 0

yit = 0 , else

uit q N(0, j 2 )

36
 , P (y = 1|xit , zt ) = P (yitW > 0|xit , zt ) (33)
µ ¶
uit (qxit + zt )
= P > (34)
j j
µ ¶
q 
= x xit + zt (35)
j j

If

xit q q N(µt , j 2xq ) (36)

µ ¶
j 2xq
, xt q q N µt , (37)
n

, p lim (xt q) = µt (38)

n<"

e and 
Then there exists parameters q, e both identifiable functions of q
and 
such that
j j

for a large N we can write:

µ ¶ ³ ´
q  e + zt 
Exit x xit + zt ' x xt q e (39)
j j

This is easily seen by rearranging eq(32);

yitW = xit q + zt  + uit = xt q + zt  + (xit  xt )q + uit (40)

By standardizing by the variance of the compound error term, we have:

yf
W e e + /it
it = xt q + zt  (41)

37
 Considering that;
3 4
(xit  xt )q + uit D µ  xt q
E (/ it |zt , xt ) = E C q = qt
j 1 + j 2xq j 1 + j 2xq

, and assuming that N is large, then by eq(38)

xt q ' µt (42)

and consequently:

/ it q N(0, 1) (43)

Considering the identity;

W
yit
yfW
it = q (44)
j 1 + j 2xq

, obviously yfW W
it > 0 if and only if yit > 0, which in turn implies;

P (yit = 1|zt , xt ) = P (yitW > 0|zt , xt ) = P (yf

W
it > 0|zt , xt ) (45)
³ ³ ´´ ³ ´
= P /it >  xt q e + zt 
e = x xt q e + zt 
e

Note the digerence between the expressions P (yit = 1|xit , zt ) and P (yit = 1|zt , xt ).
The former is the probability that individual i will give positive response at time t as
a function of both the macro variable and value of the individual characteristic, where
as the latter refers to the same probability as a function of the macro variable and the
mean value of the characteristic. The latter therefore has the interpretation of being the

38
expected proportion of positive responses in the population at time t, that is;

1 X
N
P (yit = 1|zt , µt ) = Exit (P (yit = 1|xit , zt )) ' P (yit = 1|xit , zt ) (46)
N i=1

For xt to be dependent of macro conditions, some form of non-stationary must be

assumed for the distribution of q 0 xit . In the framework sketched above, the mean value µ
is assumed to be time dependent, whereas the variance is assumed to be constant. In other
words; changes in over all credit risk is assumed only to shift the mean of the distribution
of q 0 xit . With reference to elementary finance theory, market portfolios are rarely assumed
to have a constant variance, and it is not obvious that portfolios exposed to credit risk
will be much digerent in nature. Digerent industries will depend in digerent ways on the
macro environment, and thus a macro shock, a policy shock, or an overall change in the
business cycle is expected to inflate the variance of q 0 xit . Some egort should therefore be
devoted to explore the robustness of the model to deviations from the constant variance
assumption. It is however not likely that this aspect can be modeled directly, due to
the limited time dimension of the panel in hand. Some remedy for this problem should
however be brought about if we aggregate by industry. If the aggregation is suciently
”fine-meshed”, heterogeneity in macro dependency over industries could at least partially
be considered taken care of.
So far I have sketched a way to start with the estimate of ( qj , 
j
, j 2xq ) derived from
the micro-level model, and arrive at an aggregate model for mean predicted outcome. By
conditioning on future developments of x and z the model can be used for evaluating
scenarios. Because z is likely to be policy dependent, the model could appear useful for
policy analysis. Because the panel is sub ject to limited time variation, we are however

not likely to obtain any sensible estimate of j
. We should therefore seek to find some
analogy to the case of the linear regression model sketched in section 6.1. The question
is whether we can estimate the model;

³ ´
e + zt 
Pt = x xt q e , t = 1, 2, ..., T (47)

39
, and identify of the j -coecient by some function of the estimated je -coecient.

The log likelihood for the problem is;

Xn h ³ ´i h ³ ´io
l= e + zt 
y t ln x xt q e + zt 
e + (1  y t ) ln 1  x xt q e · nt (48)
tMT

where y t and nt is the fraction of firms going bankrupt and the number18 of observations
at time t respectively.
Alternatively we can turn to minimum chi-square methods. By inverting eq(47), we
have;

e + zt 
x31 (Pt ) = xt q e (49)

, and thus we could estimate the equation;

e + zt e
x31 (y t ) = xt q  + #t (50)

as a linear regression model. Expanding x31 (y t ) around Pt by a first order Taylor ap-
proximation we have:

Yx31 (Pt )
x31 (y t ) = x31 (Pt ) + (y t  Pt )
YPt
y t  Pt
= x31 (Pt ) + (51)
)(Pt )

µ ¶
y t  Pt
E (# t ) = E =0 (52)
)(Pt )
Pt (1  Pt ) 2
V ar(# t ) = 2  wt (53)
nt ) (Pt )

,where )(Pt ) denotes the normal density function. Thus we should apply the method

18
This number will be large, and furthermore vary little over the estimation period. The eect of this
weighting should therefore be marginal.

40
 bt for wt in applying an iterative procedure.
of weighted least squares, substituting w

6.3 Identification of the micro model coecients

e parameter, we wish to derive

After we have obtained a consistent estimate of the 
the parameter  for inclusion in the micro-level model. Under the constant variance
assumption, this appears to be rather straight forward. Considering that


e
 j
= q
j 1 + j 2xq
  q
e
, = · 1 + j 2xq (54)
j j

all that is needed for identification is the estimate of j 2xq from the micro-level model.

41
7 Bibliography

Altman, E.I.(1968):”Financial Ratios, Discriminant Analysis and the Prediction of Cor-

porate bankruptcy”,Journal of Finance, September,p.589-609
Altman, E.I.,Y. Chen and Weston F.(1995):”Financial Distress and Restructuring
Models”,Financial Management,Vol.24,No.2,Summer,1995, p57-75
Beaver, W.(1966):”Financial Ratios as Predictors of Failure”,Journal of Accounting
Research,p.77-111
Bullow, J.I. and J.B. Showen(1978):”The Bankruptcy Decision” ,Bel l Journal of Econ
omics,Autumn,p.4456
Dixit, A.K. and R.S Pindyk(1993): Investment under Uncertainty, Princeton Univer-
sity Press, Princeton
Guilkey, D.K. and J.L. Murphy(1993):”Estimating and testing in the random egects
probit model”,Journal of Econometrics ,Vol.59,p.301-317
Hausman, J.A.,J. Abrevaya and Scott-Morton, F.M.(1998):”Misclassification of the
dependent variable in a discrete-response setting”,Journal of Econometrics,Vol.87,p239-
269
Laitinen, E.K. and T. Laitinen(2000):”Bankruptcy prediction. Application of the Tay-
lor’s expansion in logistic regression”,International Review of Financial Analysis.Vol.9,p372-
349
Maddala, G.S.(1983): Limited dependent and qualitative variables in econometrics,
Cambridge University Press, Cambridge
Maddala, G.S.(1987):”Limited Dependent Variable Models Using Panel Data”, The
Journal of Human Resources..Vol.3,p.307-337
Morris, R.(1997): Early Warning Indicators of Corporate Failure, Ashgate Publishing
Ltd. Hants
Myers, S.C.(1977):”Determinants of Corporate Borrowing”,Journal of Financial Eco-
nomics, November,147-175

42
 Ohlson, J.(1980):”Financial ratios and the probabilistic prediction of bankruptcy”,
Journal of Accounting Research”Spring, p.109-131
Pesola,J (2000): ”Macro Indicators for Banking Sector Stability in Finland”, Bank of
Finland , Working paper No. 9.

Wilcox J.W.(1976):”The Gambler’s ruin approach to business risk”, Sloan Manage-

ment Review,Fall,p.33-46

43
8 Appendix

8.1 Summary statistics

Table 1. Variables for which transformations are estimated

44
 Variable Frequency
a1 .0553339
a2 .0773861
Variable Mean Std.dev Min Max a3 .0793676
size 2.063386 3.804086 2.05e-10 121.8368 a4 .0755025
meanlev .429475 .0926014 .2710669 .6487265 a5 .0714291
meaneka .6449002 .0606694 .4502819 .8484569 a6 .0663525
sdtkr .2827032 .0454702 .136976 .4127187 a7 .0603352
Table2:Size and industry characteristics. a8 .052931
div .335941
taptek .2892179
Table3.Dummy variables. Fractions.

45
8.2 Regression output. Loan losses and risk weighted debt.

Regression of risk weighted debt and the change in the real price of housing index on
loan losses for the bank sector

46
Eivind Bernhardsen: A Model of Bankruptcy Prediction Working Paper 2001/10

Non-linear estimation
KEYWORDS:

Logit analysis

25125
Aggregation
Bankruptcy