Yulia Demyanyk – Iftekhar Hasan
Financial crises and bank failures:
a review of prediction methods
Bank of Finland Research
Discussion Papers
35 • 2009


E-mail: [email protected]
Bank of Finland Research
Discussion Papers
35
•
2009

Yuliya Demyanyk* – Iftekhar Hasan**

Financial crises and bank
failures: a review of prediction
methods
The views expressed in this paper are those of the authors and
do not necessarily reflect the views of the Bank of Finland.

* Federal Reserve Bank of Cleveland.
Email: [email protected].
** Rensselaer Polytechnic Institute and Bank of Finland.
Email: [email protected]. Corresponding author.

We thank Kent Cherny for excellent comments and Qiang Wu
for research assistance.

http://www.bof.fi

ISBN 978-952-462-564-7
ISSN 0785-3572
(print)

ISBN 978-952-462-565-4
ISSN 1456-6184
(online)

Helsinki 2009

3
Financial crises and bank failures:
a review of prediction methods
Bank of Finland Research
Discussion Papers 35/2009

politiikkavaihtoehtoja, joilla näitä kriisejä hallitaan. Tässä tutkimuksessa tehdään
myös yhteenveto rahoitus- ja pankkikriisien empiirisissä tutkimuksissa käytetyistä
menetelmistä. Työssä tarkastellaan lisäksi Yhdysvaltain asuntoluottojärjestelmän
kriisiin ja globaaliin rahoitusmarkkinoiden myllerrykseen liittyviä rahoitusjärjes-
telmän ja talouden piirteitä, jotka johtivat vakavaan kriisiin monessa maassa.
Tämän tutkimuksen keskeinen tarkoitus on edistää tulevaisuudessa tehtävää
empiiristä tutkimusta, jonka avulla rahoitus- ja pankkikriisien syntyä voitaisiin
estää.

Avainsanat: rahoituskriisit, pankkikonkurssit, operaatiotutkimus, varhaisten häly-
tysten menetelmät, ennakoivat indikaattorit, subprime-asuntoluotot

JEL-luokittelu: C44, C45, C53, G01, G21

5
Contents
Abstract 3
Tiivistelmä (abstract in Finnish) 4

1 Introduction 7

2 Review of econometric analyses of the subprime crisis 9
2.1 Collapse of the US subprime mortgage market 9
2.2 The subprime crisis is not unique 11
2.3 Selected analyses of bank failure prediction 12
2.4 Remedies for financial crises 13

3 Review of operations research models 16

4 Concluding remarks 24

connectionist approac h, w here network units are connected b y a flow of
information. T he structure of the model changes based on external or in tern al
inform a tion that flo w s through the n et work during the learning phase.
Com pa red to statistical methods, N N have two advantages. T h e most
importantoftheseisthatthemodelsmakenoassumptionsaboutthestatistical
distribution or properties of the data, and therefore tend to be more useful
in practical situations ( as most financial d a ta d o not meet th e s tatistical
requirem ents o f certain statistical models). A noth er advantage of the NN
method is its reliance on n on linear a ppr oaches, so that on e c an be mo re
accurate whe n testing complex data pa tterns. The nonlinearity feature o f
NN models is important because one can argue t hat the relati on bet ween
explanatory factors and the likelihood of d efault is nonlinear (several statistical
meth odologies, howev e r, are a lso able to deal with nonlinear relationsh ips
between factors in th e data).
This paper is related to w ork of Demirguc-Kun t and D etragiache (2005)
who review two ear ly w ar ning methods — signals approac h and the multivariate
proba bility model — tha t ar e frequently used in emp irical resea rch analyzing
banking c rises. Bell a nd Pain (xxxx) review the u sefuln ess a n d a p plicab ility
of the leading indicator m odels used in the empirical research analyzing and
predicting financial c rises. The authors n ote that the models need to be
impr oved in order to be a m or e useful tool for policymakers and analysts.
In this review we show that statistical tec hniques are frequently
accom p an ied by in telligenc e tec h niqu es for better model performance in the
empirical literature aimin g to better predict and a na lyze defaults and crises.
1
Chen and Shih (2006) and Boyacioglu et al (2008).
7
In most of the cases review ed , models th at use operations resea rch tech niques
alone or in combination with statistical methods predict failures better than
statistical m odels alone . In fact, hy br id intelligence sys tem s , which combine

securitized mortgage d eb t ($ 6.8 trillion).
5
In other words, as of the second
quarter of 2008, the subp rim e securitized market was roughly one-third of
the total securitized ma rket in the United States, or approximately 16 per
cent of the en tire U S m ortg ag e debt. Before the crisis, it was believ e d that a
mark et of suc h small size ( relatively to the US total mortgage market) could
not cause significan t problems outside the subprime sphere e ven i f it were
to cr ash completely. However, we no w see a sev ere ongoing crisis — a crisis
that has affected th e real economies of many countries in the world, causing
recessions, b ankin g and financial crises, a nd a global credit crunch.
The large effect of the relatively small subprime com ponen t of the mortgage
mar ket and its collapse was most likely due to the complexit y of the mar ket for
the securities that wer e created b ased on subprime m ortgages. The securities
were c rea ted by pooling individua l sub prime m ortga ges t oge ther; in addition,
2
See Demyan y k and VanHemert (2008) and Demyanyk (2008) for a more detailed
description and discussion.
3
As the total value of subprime securities issued between 2000 and 2007, calculated by
Inside Mortgage Finance, 2008.
4
Total value of mortgages outstanding in 2Q 2008. Source: Inside Mortgage Finance,
2008
5
Total value of mortgage securities outstanding in 2Q 2008. Source: Inside Mortgage
Finance, 2008
8
the securities themselves w ere again repac kaged and tranc hed to create m o re
com plica ted financial instrumen ts.

home sales/purc hases occurred), and overall econom ic slowd own cr eated a
self-sustaining loop, escape from w hich was beyond the capacity of mark et
forces to find.
Demyanyk and Van Hemert (2008) analyzed the s ubp rime crisis em pirically,
utilizing a d ur atio n statistical model t h at allo w s estimating the s o -calle d
survival time of m o rtgage loan s, ie, ho w long a loa n is expected to be
current before the very first delinquency (m issed paym ent) or defau lt occurs,
conditional on never having been d elinquent o r in default before. The model
9
also allows con tro lling fo r vario u s individu a l loan an d borrower cha rac teristics,
as w e ll as macroeconomic circumstances. According to t he estimated results,
credit score, the cumulativ e loan-to-value ratio, the mor tgage rate, and
the house pri ce appreciation hav e the largest (in absolute terms) marginal
effects and are the most important for explaining cross-sectional differences in
subprime loan performance. H owever, according to the same estima ted m odel,
the crisis in the subprim e m ortgag e m arket did not occur because housing
prices in the United States started declin ing , as many have conjectured . The
crisis had been brew ing for at least six consecutiv e years before signs of it
became visible.
The quality of subprim e mortgages h ad been deterior atin g monotonically
ev ery year since at least 2001; this pattern was m asked, however, by house
price appreciation. In other w or ds, the quality of loans did n ot suddenly
become muc h worse j ust before th e defaults occurred — the qu ality wa s poor
and w orsen ing eve ry y ear. We were able to observ e this inferior qualit y only
when the housing market started slo wing down — when bad loans could not
hide behind high house a ppreciation, and when bad loans could no longer be
refinanced.
Demy anyk and Van Hemert also sho w that the abov e-mentioned
mon otonic deterio ration o f subp rime mortgages was a (subprim e) m a rket-wide
phenomenon. They split their sample o f a ll subprime mortgages in to t he

results suggest that securitization d oes ad versely affect the screenin g incen tives
of lenders.
Mian and S ufi (2008) sh ow th at securitization is associated w ith increased
subprime lending and subsequen t defaults. M ore specifically, the au thors show
that geographical areas (in this case, zip codes) with m ore borro wers who had
credit application rejections a decade before the crisis (in 1996) had more
mortgage defaults in 2006 and 2007. Mian and Sufi also find that ‘prior to the
default crisis, these subprime zip codes (had experienced) an unprecedented
relative gro w th in m o rtga ge cred it’. The expansio n in mortgag e cred it in th ese
neighborhoods w a s combined with de clin ing inc om e growth (relative to oth er
areas) and an increase in securitization of subprim e mortgages.
Taylor (2008) blames ‘too easy’ monetary policy decisions, and the resulting
lo w interest rates bet ween 200 2 and 2004 for causing the monetary excess,
which in turn led to the housing boom an d its subsequent collapse. He
compa res th e h o using market boom that could have r esulted in the US economy
if mon etary policy had been conducted according to t he historically followed
Taylor rule — a rule that su ggested m u ch higher interest rates for the period
— with the a ctu al housin g boom . Based o n th e com pa rison, there would have
been almost no housing boom w ith the higher rates. No boom would ha ve
meant no subsequen t bust. The author dismisses the popular h y pothesis of an
excess of world savin gs — a ‘savings g lut’ — t ha t m any use to justify t he low
in terest rates i n th e econ omy, and shows that th ere was, in fa ct, a global savings
shortage, not an excess. Also, com pa ring monetary policy i n other countries
with that in the United Sta tes, Taylor notices that the housing booms were
largest in coun tries w h ere deviations of the actual interest rates from tho se
suggested by the Taylor rule were the largest.
There is a large literature that analyzes mortgage defaults. The analysis is
importan t for understanding the subprime mortgage crisis, which was triggered
by a massiv e w ave of mortgage delinquencies and f oreclosu res. Importan t
contributions to this l iterature include Deng (1997), Ambrose a nd Capone

(in 2007); Argen tina (in 1980); Chile (in 1 982); Sw eden, Norway, and Finland
in (1992); Mexico (in 1994); and T hailand, Indonesia, and Korea (in 1997) all
experienced the culmination of sim ilar (lending) boom -bust scenarios, but in
very different economic circumstances.
Reinhart and Rogoff (2008), who analyzed macro indicators in the United
States preceding the financial crisis of 2008 and 18 other post-Wo rld War II
banking crises in industrial countries, also found striking sim ilarities a m o ng a ll
of them. In par ticular, the c ou ntries experiencing the crises seem to share a
simila rity in the sign ificant increases in ho using p rices before the fina ncial crises
commenced. Ev en more striking is evidence that the United States had a much
higher gro w th rate in its house prices than the s o-called Big F ive countries in
their crises (Spain in 1977, Norway in 1987, Finland in 1991, S weden in 1991,
and Japan in 1992). In comparing the real rates of growth in equit y mark et
price indexes, the authors again find that pre-cr isis similarities are eviden t
amo ng all th e crisis co untries. Also, in comparing the curren t account as a
percen tage of gross domestic product (GD P), not only are there similarities
between coun tries, b ut the United States had la rger deficits than those of th e
other countries before their crises, reaching more than six percent of G D P.
The a uthors no ted, how ever, that there are great un certainty associated with
the still ong oin g 2008—200 9 crisis in the Un ited States; therefore, it is not
possible to project the path o f crisis resolution b ased on the experiences of
other countries.
2.3 Selected analyses of bank failure prediction
Dem irguc-K unt and Detragiache (1998) study the determ inants of the
proba bility of a b an king crisis aro un d the world i n 1980—1994 using a
multivariate Logit model. They find that bank c rises are m ore lik ely in
coun tries with low GDP gro wth, high real interest rates, high inflation rates,
and explicit deposit insurance system. Countries that are more suscep tible
12
to balance of p ayments cris es also have a higher probability of e x periencing

efficiency between 1985 and 1994. The authors find that these instan ces a re
in terrelated and G r anger -cause each other.
2.4 Rem edies for financial crises
Caprio et al (2008) indicate that recent financia l crises often occur because o f
booms in macroeconom ic sectors; the crises are revealed follo w ing ‘iden tifiable
shocks’ t h at e nd the booms. Importantly, t he u n der ly in g distortions
of economic markets build up for a long time before the crisis is identified
(Demya nyk and Van Hemert (2008 ) identify suc h a p rocess for the US sub pr ime
mor tgage crisis). Cap rio et al (2008) discuss the ro le o f financial deregulation
in pr edictin g crises and iden tify a mec h anism for interaction between the
governm ents and regulated institutions. The authors propose a series o f
reform s that could prevent future crises, such as lending reform, rating agency
reform and securitization reform. M o st importantly, a cco rding to the authors,
13
regulation and supervision should be re-strengthened to prevent such crises in
the future.
In his researc h, Hunter (2008) attempts to understand the causes of, and
solutions for, the financial crises. He defines the beginning o f th e recent crisis
in the United States to be the poin t in time when inter-bank lending stopped
in the Feder al Fu n ds M arket. Following this definition, the US crisis began
around October 8, 2008, when the Feder al Funds Rate hit a high of seven
percen t during in traday trading. According to Hunter, the primary reason
for trading h alt was that ban ks were unsure about the exposure of their
counterparties to MBS ris k: ‘If a b a nk ha s a large share of its asset portfolio
devoted to MB S , then selling MB S to g et operatin g cash is infeasible when
the price of M BS has declined significantly. Banks in this situa tion are on th e
brink of insolvency and may indeed have difficulty r epaying loan s they receive
through the Federal Funds Market’. The autho r suggests several solutions to
the crisis. Among them, he emphasizes the importance of transparency in the
operation of and analysis by MBS insurers and bond rating agencies. He also

hold them un til they are repaid, to an ‘originate-to-distribute’ banking m odel,
in which loans are pooled, tranched and then sold via securitization. This
14
transform ation can red uce banks’ monito ring incentives and increase their
possibility of if th ey hold a large amou nt of such securities withou t f ully
understand ing the associated credit r isk.
Brunnermeier further i den tifies several econ om ic mec hanisms throug h
which the mortgage crisis w as am plified into a broader financial crisis. All
of the mec han isms begin with the drop in house prices, wh ich eroded the
capital of fin an cia l in stitu tion s. At the s a m e time, lenders t ig htened lend ing
standards and margin s, whic h cau sed fire sales, further pushing down prices
and tighten ing credit supplies. When ba nks became concerned about their
abilit y to access capital markets, they began to hoard funds. Consequ ently,
with the dro p in b alance sheet capital and difficulties in accessin g additional
funding, b anks that held large am o unts of MB S failed ( eg, Bear Stearns,
Lehman Brothers, and Washington Mutual), causing a sudden shock to the
financial mark et.
Sev eral researc hers conclude that the ongoing crisis does not reflect a failure
of free mark ets, but a rather reaction of market participants to d istorted
incentives (Demirgu c-K unt and Serv en, 2009). Dem irguc-Kunt and Serv en
argue that the ‘sacred cows’ of financial and macro policies are not ‘dead’
because of the crisis. Man ag ing a s ys tem ic panic requires policy decisions t o be
made in differen t stages: the im m ediate conta inment stage and a longer-term
resolution accompan ied by structural reform s. P olicies emplo yed to r eestablish
confidence in the short ter m , such as pr oviding blanket gua rantees or
go vernm ent buying large stakes in the financial sector, are fraught with moral
hazard problem s in the long term and might be in terp reted as perman ent
deviation s from well-e stablishe d polic y positions by the market. Th e long-term
fin an cial sector policies should align private incentives with public inte rest
without taxing or subsidizing private risk-taking (Demirg uc-K u nt and Serv en,

3 R eview of o peratio ns researc h m odels
In this section, w e describe selected operation s research models that are
frequently used in the em pirical literature to predict defaults or failures of
banks and that could be used to predict defaults of loans or non-financial
institution s.
Predicting the defau lt risk for banks, loans and securities is a classic, yet
timely issue. Since the work of Altman ( 1968), who s uggested using the
so-called ‘Z score’ to predict firms’ default risk, hundreds of research articles
have studied this issue (for reference, see t wo review articles: K u ma r and Ravi
(2007) and Fethi and Pasiouras (2009)).
Sev e ral studies have shown that intelligence modeling tec h niques used in
operations research can be applied for predicting the bank failures and crises.
For example, Celik and Karatepe (2007) find that artificial neural network
models can be used to forecast t h e rates of n on-performing loans relative
to total loans, ca pital relativ e to assets, profit relativ e to assets, and equity
relative to assets. In another example, Alam et al (2000) demonstrate t hat
fuzzy clustering and self-organizing neural net works provide classification tools
for id entifying potentially failing ban k s.
Most cen tral banks h ave emplo yed va rious Early Warning Systems (EWS)
to monitor the risk of banks for y ea rs. Ho wev er, th e repeated occurrence
of ba nkin g crises during the past two decades — such as the Asian crisis, the
Russia n bank crisis, and the Brazilian bank crisis — indicates that s afegua rdin g
the banking system is no easy task. According t o the Federal D eposit Insurance
Corporation Impro v ement A ct of 1991, regulators in the United States m ust
conduct on-site examinations of bank risk ev ery 12—18 months. Regulators use
a rating system (the CAM E L S rating) to indicate the safety and soundness
of banks. CAME LS ratings include six parts: capital adequacy, asset qualit y,
man agem ent expertise, earnings strength, liquidity and s ensitivity to market
risk.
Da vis and Karim (2008a) evaluate statistical and intelligence techniques in

Karels and P rakash (1987), Haslem et al (1992)). There are three subcategories
of DA: Linear, Multivaria te, and Q u adra tic. One drawback of DA is tha t it
requires a normal distribution of regressors.
8
Wh en regressors are not norma lly
distributed, maximum likelih ood methods, suc h as Lo git, can be used.
9
DA is
a tool for analyzing cross-section a l data . If one needs to a nalyze time series
data on bank firm, or loan defaults, hazard or duration analysis models can
be used instead of DA models.
10
Can bas et al (2005) propose an Integrated Early Warning System (IEW S)
that combines D A, Logit, P robit, and P rincipal Component Analysis (PCA),
which can help predict ban k failure. First, th ey use PCA to detect three
financial componen ts that significantly explain the changes in the financial
condition of banks. They then employed DA, Logit and Probit regressio n
models. By combining all these together, they construct an IEW S. The authors
use the data for 40 privately o wned Turkish commercial banks to test t he
predictive power of t he IEW S, concluding that the IEWS has more predictive
ability than the other models used in the literature.
Am ong in telligence tec hniques, Neural Networks (N N ) is the most widely
used. The NN model ha ve developed out of the fields of artificial in te llig en ce
and b rain modeling, and conta ins m a thematical an d alg orith m ic elements that
mimic t h e b iological neural networks of the hum an nervous system . The
meth od consid ers an i nterrelated g rou p of artificial n euro ns and processes
inform ation associated with them using a so-called connectionist approach,
8
Martin (1977) is an early study that uses both Logit and DA statistical methods to
predict bank failures in the period from 1975 to 1976, based on data obtained from t he

comm only used methods f or classifica tion and prediction problems. Many
studies com p are the classification and pred iction accuracy between BPN N and
other m ethods and find that, in most cases, B PN N outperform s other models.
For exam p le, Tam (1991) uses a BPNN model to predict bank failu res in
a sam ple o f Texas ban ks o ne year and tw o years prior to their failures. The
input v ariables he uses are bases on the CAM ELS criteria. He finds that BPNN
outperforms all oth er m ethods, such a s DA, L ogit, and K -near est neig hbor (this
meth od is described below) in terms of their pred ictive accur acy. Similarly,
sev e ral other studies, briefly described below, find that BPNN offers a better
prediction or a better classification accuracy than other methods.
Ravi and Pram odh (2008) propose a Principal Com ponent N eural Network
(PCNN) architecture for bankruptcy predic tion in commercial banks. In this
architectu re, the hidden layer is comp letely replaced by what is referred to as
a ‘princip al com ponen t layer’. This layer consists of a few selected components
that perform the function of hidden nodes. The authors tested the framew ork
on a d ata from Spanish and Tu rkish banks. According to the estimated results,
hybrid models that combine PCNN and se veral othe r models predict b an kin g
bankrup tcy outperform other c lassifiers used in the litera tu re.
Tam a nd Kiang (1992) compare the po wer of linear discriminant analysis
(LDA), Logit, K-Nearest Neighbor (described below ), Intera ctive Dic hotom izer
3 ( ID3), feedforw ard NN and BPNN on bank failure prediction problems.
They find that BPNN outperforms the other techniques for a one-y ear-prior
training sample, while DA outperforms the oth ers for a tw o-years-prior tra ining
sample. Ho wev er, f o r h old out s amples, BPN N outperforms t he others in
both th e one-y ear -prior and the two-years-prior samples. In the jac kknife
meth od, BPN N a lso o utperforms o thers in both the one-year-prior and the
18
t wo-y ears-p rior h oldou t sam ples. In all, they conclude that NN ou tperforms
the DA method.
Bell (1997) compares Logit and BPNN models in predicting bank failures.

theadequateselectionofcutpointsforeachofthevariables,sothatallfailed
banks can be located below some threshold and all non-failed banks above it.
Kolari e t al (2002) develop an EW S based on Logit a nd the Trait
Recognition method fo r large U S banks. The Logit model correctly classifies
o ver 96% of the banks one y ear prior to failure and 95% of the banks tw o years
prior to failure. For the Trait Recog nitio n model, half of the original sample is
used. They find th at with data classification both one year and two y ea rs p rior
to failure, the accuracy of the Trait R eco gnition model is 100% . T herefore,
they conclu de th at the Trait Re cogn ition m odel outperforms the Logit model
in term s of type-I and type-II errors.
Lanine and Vander Vennet (2006) e m ploy a Logit model and a Tr ait
Recognition app roa ch to predict failures among Russian com m ercial banks.
The authors test the predictive power of the t wo models based on their
prediction accuracy using holdout samples. Although both models perform
better than the benchmark, the Trait Recognition approach outperforms Logit
19
in both the original and the holdout samples. For the predictable variables,
they find that expected liquidity plays an im portan t role in bank failure
prediction, as well as asset q uality a nd capital adequacy.
TheSupportVectorMachine(SVM)techniqueisbasedontheStructural
Risk Minim ization (SRM) principle fro m com p utatio nal learning theory, which
w as introduced by Vapnik (1995). In the SVM method, input data is structured
as t wo sets of vectors in a m u lti-dim ensio na l space. T h e purpose is to m a ximize
the m argin between the two data sets. In order to calculate the margin, t wo
parallel h yperplanes need to be constructed, one on each side of the separating
h yperplane, which are forced against the tw o data sets. A good separation
can be a chiev ed by the hyperplane that has the largest distance f rom the
neighboring data poin ts of both classes; th e lar ger the m a rgin, the better the
generalization error of the classifier. In sum , SVM uses a special linear model
and the optimal separating h yperplane to ac hieve the maxim um separation

and patterns in landscape, s pam classification,oranyotherdistributionof
objects and events. One can determine if objects or events are random ,
clustered, or distributed r egularly. T he K-nearest n eigh bor (K -NN) is a
modified Nearest Neighbor technique. I n this model, K is a positive, usually
20
small, integer. An object (for example, a b a nk) is assig ned to th e c lass
most commo n amongst its K nearest neighbors ( th e c la ss is either ‘failed ’ or
‘non-fa iled’).
Zhao et al (2009) compare the performance of several factors that are used
for predicting bank failures b ased on Logit, D T , NN, a nd K-NN models. The
authors find that a m odel choice is important in terms of explanatory power
of predictors.
The S o ft C ompu ting t echnique is a hybrid syste m combining intelligence
and sta tistical techniques. Specifically, it refers to a combination o f
computational tec hniques in order to model and analyze complex phenomena.
Com pa red to traditio nal ‘hard’ comp uting techniques — whic h use exact
computations and algorithms — soft computing is based on inexact
compu tation, tr ial-and-error reasoning, and subjective decision making. Such
compu tation builds on m ath em a tical formalization of the cognitiv e processes
similar to th ose of human minds. More informat ion is available in Back an d
Sere (1996), Jo and Han (1996), Tu ng et al (2004).
Data Envelop m ent Analysis ( D EA) is a non-param etric perform an ce
meth od used to m easu re the relative efficiencies of organizational or
decision-making units (DMUs). DEA applies linear programming to observing
inputs consumed and outputs produced by decision-making u nits (suc h as
branches of a bank or departments of an institution). It constructs an efficien t
production frontier based o n best observed practices. Eac h DM U ’s efficiency
is then measured against this c omputed fron tier. The relativ e efficiency is
calculated by obt aining the ratio of the weighted sum o f a ll outputs and
the weighted sum of all inputs. The weights are selected to ac hieve Pareto

scores, t he authors use a Ba yesian approac h to the problem set up around
sim ulation techniques. They also test the new m ethods on the efficiency of
Gr eek banks, and find that the m ajority of t he Greek banks operate c lose to
mar ket best-practices.
Cielen et al (2004) com p are the performance of a DE A model, Minim ized
Sum of Deviat ions (MSD ), and a rule induction (C5.0) model in ba nkr up tcy
prediction. M SD is a com bination of linear programming (LP) and D A. Using
data from the National B ank of Belgium, they find that MSD, DEA and C5.0
obtain the co rrect classification rates o f failure for 78.9%, 86.4% and 85.5%
of ba nk s, respectively. They c on clude that DE A outperfor m ed the C5.0 and
MSD models in terms of accuracy.
Kosmidou and Z opounidis (2008) develop a bank failure prediction
model based on a m ulticriteria decision tec h niqu e called UT ilites Additives
DIScriminan ts (UTA DIS). The purpose of UTADIS method is to dev elop a
classification model throu gh an additive value fun ction. Based on the values
obtained from the additiv e value f unction, t he authors classify banks in to
multiple groups by compa ring t h em with s om e r eferen ce profiles (also called
cut-off poin ts). UTAD IS is well suited to the o rdinal classification problem s
and it is not sensitive to th e statistical proble m s because the additiv e utilit y
function is performed through m athematical linear programming tec hniques
instead of statistical methods. Using a sample of US bank s for the years
1993—2003, the authors use this tec hnique to differentiate US banks bet ween
failed and non-failed. The results show that UTA D IS is quite efficient
for the evaluation of bank failure as early as four y ears before it occurs.
The authors also compare UTAD IS with other traditional multivariate data
analysis techniques and find that UTADIS performs better, and could be used
efficiently for predicting bank failures.
The M u lticriteria D ecision Aid (M C DA) meth od is a model that allow s
for t h e analysis of several preference c riter ia simultaneously. Zopounidis
and Doumpos (1999b) a pply MCDA to sorti ng problems, where a set of

results show that ‘equity/customer and short-term funding, net interest margin
and return on a v erage equit y, are the most important financial variables.
The n umber of shareholders, the n umber of subsidiaries and the banking
en vironment of the coun try’ are the most im portant non-financial factors. T he
authors compare the accuracy of this prediction m odel with that of D A and
ordered Logit; they find that MCDA is more efficient and that it replicates the
Fitch credit ratings w ith the ‘satisfactory accuracy’.
Niem ira and Saat y (2004) use a m ultip le criteria decision-making model to
pre d ict the likelih ood o f a financial crisis based on an A n alytic N etwork Process
(ANP) framew ork. They test the m odel for the US bank crisis during 1990s,
and find that the ANP analysis provides a structure that can reduce judgmental
forecast error through im p roved reliab ility of information pr ocessing. They
conclude that the ANP framework is more flexible and is m ore comprehen sive
than traditional models, and it is a promising methodology to forecast the
proba bility of c rises.
Ng et al (2008) propose a Fu zzy Cerebellar Model Articu lation
Controller model (FCMAC) based on a compositional rule of inference called
FC M AC-CRI(S). T h e new arc hitecture integrates fuzzy systems and N N to
create a h ybrid structure called neu ral fuzzy networks. This new network
operates t hrough localized learning. I t tak es as inputs data f rom public
financial information and analyzes p atterns of financial d istress th ro ugh fuzzy
IF-TH E N r ules. Su ch processing can provide a basis for an EWS and
insights for various aspects of fina ncial distress. T he authors compar e the
accuracy of F CM A C-CRI(S) to Cox’s proportional hazard model and the
GenSoFNN-CRI(S) netw ork model and find that the performance of t he new
approach is better than that of the benc hm ark m odels.
23