Você está na página 1de 28

European Journal of Operational Research 180 (2007) 128

www.elsevier.com/locate/ejor

Invited Review

Bankruptcy prediction in banks and rms via statistical


and intelligent techniques A review
P. Ravi Kumar, V. Ravi *

Institute for Development and Research in Banking Technology, Castle Hills Road # 1, Masab Tank, Hyderabad 500 057 (AP), India

Received 19 July 2005; accepted 11 August 2006


Available online 17 November 2006

Abstract

This paper presents a comprehensive review of the work done, during the 19682005, in the application of statistical and
intelligent techniques to solve the bankruptcy prediction problem faced by banks and rms. The review is categorized by
taking the type of technique applied to solve this problem as an important dimension. Accordingly, the papers are grouped
in the following families of techniques: (i) statistical techniques, (ii) neural networks, (iii) case-based reasoning, (iv) decision
trees, (iv) operational research, (v) evolutionary approaches, (vi) rough set based techniques, (vii) other techniques subsum-
ing fuzzy logic, support vector machine and isotonic separation and (viii) soft computing subsuming seamless hybridiza-
tion of all the above-mentioned techniques. Of particular signicance is that in each paper, the review highlights the source
of data sets, nancial ratios used, country of origin, time line of study and the comparative performance of techniques in
terms of prediction accuracy wherever available. The review also lists some important directions for future research.
 2006 Elsevier B.V. All rights reserved.

Keywords: Bankruptcy prediction; Banks; Firms; Statistics; Neural networks; Fuzzy logic; Case-based reasoning; Decision trees; Evol-
utionary approaches; Operations research; Rough sets; Support vector machine and soft computing; Intelligent techniques

1. Introduction upon (i) how nancially solvent it is at the inception,


(ii) its ability, relative exibility and eciency in cre-
The prediction of bankruptcy for nancial rms ating cash from its continuous operations, (iii) its
especially banks has been extensively researched access to capital markets and (iv) its nancial capac-
area since late 1960s [3]. Creditors, auditors, stock- ity and staying power when faced with unplanned
holders and senior management are all interested in cash short-falls. As a bank or rm becomes more
bankruptcy prediction because it aects all of them and more insolvent, it gradually enters a danger
alike [121]. The health of a bank or rm in a highly zone. Then, changes to its operations and capital
competitive business environment is dependent structure must be made in order to keep it solvent
(http://www.solvency.com/solvency.htm).
*
The most precise way of monitoring banks is by
Corresponding author. Tel.: +91 40 23534981; fax: +91 40
23535157.
conducting on-site examinations. These examina-
E-mail addresses: vravi@idrbt.ac.in, rav_padma@yahoo.com tions are conducted on a banks premises by regu-
(V. Ravi). lators every 1218 months, as mandated by the

0377-2217/$ - see front matter  2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2006.08.043
2 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

Federal Deposit Insurance Corporation Improve- tion 3, in the following aspects: (i) It is categorized
ment Act of 1991. Regulators utilize a six part rat- with respect to the techniques employed. (ii) It is
ing system to indicate the safety and soundness of by far, the most comprehensive and organized with
the institution. This rating, referred to as the CAM- respect to various dimensions of review such as the
ELS rating, evaluates banks according to their basic nancial ratios used, source of data set, country of
functional areas: capital adequacy, asset quality, origin, time line of study, type of industry viz., bank
management expertise, earnings strength, liquidity or rm as detailed in Tables 2 and 3. (iii) It also
and sensitivity to market risk. While CAMELS rat- gives the reader a brief overview of the techniques
ings clearly provide regulators with important infor- employed. (iv) It gives the reader an unbiased com-
mation, Cole and Gunther [20] reported that these parison of the statistical and intelligent techniques.
CAMELS ratings decay rapidly. Fraser [30] noted (v) It can be considered as a starting point to pursue
that banks perform better by holding relatively research in bankruptcy prediction.
more securities and fewer loans in their portfolios.
Gady [32] and Fraser [30] showed that core deposit 1.1. Earlier reviews
funding is benecial for banks, particularly demand
deposits, which are non-interest bearing. Gady [32] This section presents an overview of the review
has indicated that high-performance banks are able papers published earlier on bankruptcy prediction.
to generate more interest or non-interest income Scott [99] reviewed several empirical and theoretical
than under performing banks. Wall [117] observed models and found a substantial amount of overlap
that higher prot banks rely more on equity fund- between these two lines of research. He concluded
ing. Brewer et al. [15] observed that rms used that the success of empirical models suggested the
the derivative instruments to change their risk expo- existence of a strong underlying regularity, though
sure. They also concluded that there was a negative they are not based on explicit theory. Dimitras
correlation between risk and derivatives usage. et al. [25] presented a review of the articles published
Haslem et al. [44] determined the impact of types during 19321994 in various journals specialized in
of strategies followed by individual banks related accounting, nance, operation research and decision
to the relative protability performance. Kwast science. They selected 47 articles restricted to (i)
and Rose [60] employed statistical cost accounting journal articles that presented models and (ii) indus-
techniques to examine the relationship between trial and retail applications. Their framework of
bank protability and pricing and operating study was based on the country, year of publication,
eciency. industry type, sample periods and method. They
This paper presents a comprehensive survey of a reviewed a total of 59 models and variables used
number of research works published during 1968 in these articles. OLeary [84] proposed a meta
2005, where various statistical and intelligent tech- analysis of the use of neural networks to predict
niques were applied to solve bankruptcy prediction corporate failure. His review was across dimensions
problem in banks and rms. The usefulness of the such as the software used, the input variables, the
current review paper is that the papers are catego- nature of the hidden layer used, number hidden
rized, primarily, according to the technique nodes and statistical analysis of results. Tay and
employed therein. This aspect paves the way for Shen [110] reviewed the application of various
the researchers from the statistical and the articial rough set based models to bankruptcy prediction
intelligence/soft computing community to shift and concluded that rough set models were promis-
focus on this exciting problem of bankruptcy pre- ing alternative to conventional methods. Daubie
diction by applying their newest and novel ideas. and Meskens [23] reviewed the prediction models
This will only enhance and enrich the bankruptcy applied to business failure in banks and rms. They
prediction research further as these researchers are discussed the causes, symptoms, process and reme-
likely to come out with robust and high performing dies of failure. They reviewed papers published dur-
models. The review paper is aimed at attracting ing 19682000 in journals specialized in accounting,
fresh graduates, researchers and academics from nance, operations research and decision science.
nance, banking, statistics and articial intelli- Calderon and Cheh [17] reviewed the applications
gence/soft computing, since bankruptcy prediction of neural networks in bankruptcy prediction. For
is a multi-disciplinary area. The current review is every study they presented the number of input vari-
dierent from the earlier reviews, presented in Sec- ables, learning sample size, testing sample size and
P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128 3

software used. They explained the comparative e- ratios used as explanatory variables, (iv) the years
ciency of dierent techniques. during which data was collected, (v) number of
The rest of the paper is organized as follows: years prior to which prediction was made (if avail-
Section 2 presents an overview of the intelligent able) and (vi) results obtained (when available in
techniques. Section 3 reviews papers dealing with concise form) in each study. Further, the review
statistical techniques while Section 4 presents the concentrated on the papers published in peer-
works employing various NN architectures. Section reviewed journals/international conferences/edited
5 presents CBR applications while Section 6 volumes in the areas of accounts, nance, manage-
describes decision tree applications to the problem. ment, operational research, neural networks, expert
Section 7 reviews the applications of evolutionary systems and decision support systems. All the
approaches and Section 8 presents applications of unpublished works in terms of Ph.D. thesis, work-
operational research to the problem. Section 9 pre- ing papers and internal reports are excluded from
sents the papers employing rough sets and Section the scope of the review. Moreover, in a paper, when
10 describes the application of other techniques multiple techniques are compared in their stand-
subsuming fuzzy logic, support vector machine alone mode, the technique proposed in that paper
and isotonic separation to the problem. Finally, is taken as the main criterion and accordingly the
Section 11 presents the soft computing applications. paper is categorized in that family. For example,
Section 12 presents some of the important insights when in a paper, NN, DA and logit (logistic regres-
that emerged out of the survey. Section 13 suggests sion) are compared in their stand-alone mode and
some future research directions in the eld and con- NN is proposed in that paper, then the paper is
cludes the review. reviewed under the NN section.
The present review indicates that neural networks
2. Review methodology family with 25 papers is the most widely applied
technique. The soft computing family with 15 papers
As mentioned earlier, the review is conducted in and statistical family with 11 papers follow this.
two broad categories: (i) statistical and (ii) intelli- Then, rough sets were applied in six papers and other
gent techniques. Among statistical techniques, the techniques accounted for ve papers, while CBR and
methods covered are: linear discriminant analysis OR techniques had four papers each. Also, evolu-
(LDA), multivariate discriminate analysis (MDA), tionary approaches and decision trees gured in
quadratic discriminant analysis (QDA), logistic three and two papers, respectively. Further, ve ear-
regression (logit) and factor analysis (FA). The lier review papers are also included in the current
intelligent techniques covered in the study belong review. Table 1 presents the main idea behind the
to (i) dierent neural network (NN) architectures intelligent techniques, their advantages and disad-
including multi-layer perception (MLP), probabilis- vantages. Table 2 lists the nancial ratios used in
tic neural networks (PNN), auto-associative neural each of the papers. Table 3 presents for all papers,
network (AANN), self-organizing map (SOM), the details such as the entity for which the study
learning vector quantization (LVQ) and cascade was done banks or rms; source of data; country
correlation neural network (Cascor), (ii) decision of origin; sample size; techniques used in the study
trees, (iii) case-based reasoning, (iv) evolutionary and the time line of the study and the number of
approaches, (v) rough sets, (vi) soft computing years prior to prediction, wherever available.
(hybrid intelligent systems), (vii) operational
research techniques including linear programming 2.1. Overview of intelligent techniques
(LP), data envelopment analysis (DEA) and qua-
dratic programming (QP), (viii) other intelligent 2.1.1. Fuzzy set theory
techniques including support vector machine, fuzzy Fuzzy set theory, proposed by Zadeh [124], has
logic techniques. Under each category, the papers found a number of applications. It is a theory of
are reviewed in the chronological order. Thus the graded concepts. It provides a mathematical frame-
most important dimension of the present review is work where vague, conceptual phenomena can be
the type of techniques applied. The review is con- rigorously studied [127]. Fuzzy logic models human
ducted across other dimensions also such as (i) experiential knowledge in any domain. When
study pertaining to bank or rm (type of rm if applied to solve process control or prediction prob-
mentioned), (ii) source of data, (iii) various nancial lems fuzzy logic takes the help of the knowledge of
4 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

Table 1
Merits and demerits of intelligent techniques
Technology Basic idea Advantages Disadvantages
1 FL Models imprecision and ambiguity Good at deriving human Arbitrary choice of Membership
in the data using fuzzy sets and comprehensible fuzzy ifthen rules; function skews the results, although
incorporates the human It has low computational triangular shape is the most often
experiential knowledge into the requirements used one. Secondly, the plethora of
model choices for membership function
shapes, connectives for fuzzy sets
and defuzzication operators are the
disadvantages
2 NN Learn from examples using several Good at function approximation, The determination of various para-
constructs and algorithms just like forecasting, classication, clustering meters associated with training
a human being learns new things and optimization tasks depending algorithms is not straightforward.
on the neural network architecture Many neural network architectures
need a lot of training data and
training cycles (iterations)
3 GA Mimics Darwinian principles of Good at nding global optimum of a Does take long time to converge;
evolution to solve highly nonlinear, highly nonlinear, non-convex May not yield global optimal
non-convex global optimization function without getting trapped in solution always unless it is aug-
problems local minima mented by a suitable direct search
method
4 CBR Learns from examples using the Good for small data sets and when Cannot be applied to large data sets;
euclidean distance and k-nearest the data appears as cases; similar to poor in generalization
neighbor method the human like decision-making
5 Rough sets They use lower and upper They yield ifthen rules involving It can be (a) sometimes impractical
approximation of a concept to ordinal values to perform to apply as it may lead to an empty
model uncertainty in the data classication tasks set (b) sensitive to changes in data
and (c) inaccurate
6 SVM It uses statistical learning theory to It yields global optimal solution as Selection of kernel and its
perform classication and the problem gets converted to a parameters is a tricky issue. It is
regression tasks quadratic programming problem; It abysmally slow in test phase. It has
can work well with few samples high algorithmic complexity and
requires extensive memory
7 Decision They use recursive partitioning Many of them can solve only Over tting can be a problem. Like
trees technique and measures like classication problems while CART neural networks, they too require a
entropy to induce decision trees on solves both classication and lot of data samples in order to get
a data set regression problems. They yield reliable predictions
human comprehensible binary if
then rules
8 DEA It uses linear programming to rank It has found numerous applications It yields only relative scoring of the
various alternatives/business units and gives exact solution business units and not absolute
according to some input and out ratings
variables
9 SC Hybridizes fuzzy logic, neural It amplies the advantages of the Apparently, it has no disadvantages.
networks, genetic algorithms, etc. intelligent techniques while However, it does require good
in several forms to derive the simultaneously nullifying their amount of data, which is not exactly
advantages of all of them disadvantages a disadvantage nowadays

the domain expert and employs fuzzy mathematics system of massively parallel, interconnected com-
to come out with fuzzy inference systems. Fuzzy puting units called neurons, arranged in layers.
logic can also be used to derive fuzzy ifthen rules The neural networks found extensive applications
from data to solve classication problems. in nancial services. The multi-layer perceptron
(MLP) [95], radial basis function network (RBFN)
2.1.2. Neural networks [81], probabilistic neural network (PNN) [104], cas-
Neural networks [62,72] oer a computational cade correlation neural network (Cascor) [28],
paradigm inspired by biological neural networks learning vector quantization (LVQ) [35,56], self-
of human nervous system. A neural network is a organizing feature map (SOM) [56] are some of
P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128 5

Table 2
Financial ratios (variables) in each paper
Ref # Variables
[1] Cash ow/total liabilities, current assets/current liabilities (current ratio), inventories turnover, net income/total sales, net
income/total assets, net working capital/total assets, owners equity/total assets, (total borrowings + bonds payable)/total
assets
[2] Net income/total assets (return on assets (ROA)), net loan losses/adjusted assets, net loan losses/total loan, (net loan
losses + provision for loan losses)/income, non-performing loans/total assets
[5] Common equity/total capital (capitalization), cumulative protability, debt services, stability of earnings, roa, liquidity, size
[6] Debt cost, debt quality, growth, indebtedness, share of labour costs, short-term liquidity, size, turnover of assets
[7] Book value/total assets, cash ow/total assets, gross operating income/total assets, ROA, price/cash ow, rate of change of
cash ow per share (ROC), rate of change of stock price, stock price volatility
[8] Cash/current liabilities, cash/net sales, cash/total assets, cash ow/current liabilities, cash ow/total assets, cash ow/total debt,
current ratio, current assets/net sales, current assets/total assets, current liabilities/equity, earnings before interest and taxes
(EBIT)/total interest payments, equity/xed assets, equity/net sales, inventory/net sales, long-term debt/equity, market value of
equity/book value of debt, net income/total assets, net quick assets/inventory, net sales/total assets, operating income/total
assets, quick assets/current liabilities, quick assets/net sales, quick assets/total assets, rate of return/common stock holders,
retained earnings/total assets, return on stock, total debt/equity, total debt/total assets, working capital/equity, working
capital/net sales, working capital/total assets
[9] EBIT/total assets, market capitalization/total debt, retained earnings/total assets, sales/total assets, working capital/total assets
[11] Complexity of capital structure, degree of competitiveness, rm age, fraud, intangible assets/net sales, natural log of total assets
deated the gross domestic product, ROA, ownership concentration, past losses, resignation, secured interest bearing debt/
total liabilities, total interest bearing debt/total liabilities
[12] Agricultural loans/total assets, commercial real estate loans/total assets, construction loans/total assets, income before extra
items, large time deposits/total assets, insiders loans over net loans, natural log(total assets), net charge os/total loans, ROA,
net interest income/total assets, net loans/total assets, non-interest income/total assets, non-performing loans/primary capital,
non-performing loans/total assets, past due loans/gross loans, primary capital/adjusted assets, provision for loan losses/total
assets, restructured loans/gross loans, return on equity, security gains (losses) and extra items/total assets, short-term assets less
large liabilities/total assets, total capital/total loans, total equity capital, total overhead expenses/total assets, undivided prot
and capital reserve to total assets, yield on total assets
[13] Cash ow/total debt, current ratio, current liabilities/total debts, gross prot/sales, net income/stockholders equity, roa, sales/
total assets
[16] Cash/current liabilities, cash/total assets, cost of goods sold/inventory, current ratio, current assets/total assets, current assets/
total sales, current liabilities/total assets, EBIT/total assets, inventory/sales, net income/net worth, net income/sales, ROA,
quick assets/sales, quick assets/total assets, retained earnings/inventory, retained earnings/total assets, sales/cash, sales/net
worth, sales/total assets, total assets/gross national product (GNP) price-level, total liabilities/net worth, total liabilities/total
assets, working capital/sales
[18] Quick ratio, income ratio, interest expenses/average non-protable assets, interest expenses/average protable assets, interest
expenses/total expenses, interest income/interest expenses, liquid assets/(deposits + non-deposit funds), liquid assets/total
assets, net working capital/total assets, (salary and employee benets + reserves for retirement)/no of personnel, (shareholders
equity + total income)/(depreciation + non-depreciation funds), (shareholders equity + total income)/total assets,
(shareholders equity + total income)/(total assets + contingencies and commodities), standard capital ratio
[19] Cash/restricted current assets, equity ratio(equity/total assets), expired taxes, retained earnings/total assets, inventories, gross
return, coverage of debt, net return, current ratio, quick ratio, debt ratio
[24] Abnormal increase in inventory and receivables, quick ratio, current ratio, debt/equity, dividend, funds ow ratio, ROA, net
worth, sales/total assets, total assets, trend in net income, working capital/total assets
[26] Current ratio, current liabilities/total assets, gross prot/total assets, inventories/working capital, (long-term debt + current
liabilities)/total assets, net income/gross prot, ROA, net worth/(net worth + long-term debt), quick assets/current liabilities,
working capital/net worth
[29] Quick ratio, current ratio, income ratio (income/working capital)
[31] Cash ow/total debt, current ratio, current assets/total assets, EBIT/total assets, log(interest coverage + 15), log (total assets),
market value of equity/total capitalization, ROA, quick assets/current liabilities, quick assets/total assets
[36] Working capital/total assets, retained earnings/total assets, EBIT/total assets, market value of equity/total assets, sales/total
assets
[37] EBIT/total assets, net income/net worth, total liabilities/total assets, total liabilities/cash ow, interest expenses/sales, general
and administrative expensive/sales, managers work experience, rms market niche position, technical structure facilities,
organization-personnel, competitive advantage of rms, market exibility
[44] Domestic cash, domestic investment securities, foreign cash, foreign investment securities, net domestic loans, net foreign loans,
net income, total assets, total domestic interest bearing deposits, total domestic non-interest-bearing deposits, total equity
capital, total foreign interest bearing deposits, total foreign non-interest bearing deposits
(continued on next page)
6 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

Table 2 (continued)
Ref # Variables
[45] EBIT/total assets, market value of equity/total debt, retained earnings/total assets, sales/total assets, working capital/total
assets
[51] Net operating cash ow to total assets, cash resources to total assets, cash ow cover (net operating cash ow to annual interest
payments), sales revenue to total assets, total debt to total equity, total debt to gross operating cash ow, working capital to
total assets (wc = current assets current liabilities)
[53] Gross prot margin ((net sales-cost of goods sold)/net sales), market value of common stock, natural logarithms of tangible
asset turnover, (net income + depreciation)/number of shares, (cash ow per share), net income/total of common equity
(earnings per share), sales/cash, sales/inventories, sales/receivables, total debt/total assets, total debt/total capital, working
capital/total assets
[55] Equity ratio (equity/total assets), net income before depreciation and extraordinary items, net income before depreciation and
extraordinary items of the previous year, operating margin
[57] Allowance for loan losses/total assets, bank holding companies total assets, bank total assets/bank holding co. (bhc) total
assets, certicate of deposit/total deposits, maximum change in assets/mean assets, maximum change in assets/mean change
assets, maximum change in loans past due at least 90 days/mean of numerator, net income after taxes/total assets, net interest
income/total assets, net loan charge-os/total assets, non-deposit liabilities/total liabilities, provision for loan losses/total
assets, sum of key asset accounts/total assets, total assets, total equity/total assets, total loans and leases/total assets, total
securities/total assets
[62] EBIT/total assets, market value of equity/book value of debt, retained earnings/total assets, sales/total assets, working capital/
total assets
[64] Capital expenditure, common shares traded, consumer price index, current account balance/gross domestic product, current
assets/common shareholders equity, depreciation expenses, dividend/share, earnings/share, eective exchange rate, federal
budget/gross domestic product, government spending/gross domestic product, (long-term debt + short-term debt)/total
assets, market capitalization, money supply, net income/net sales, net sales/total assets, pre-tax income/net sales, purchase
price of crude oil, relative strength index, research expenses, short-term interest rate, spread between short-term and long-term
interest rate, tax deferral and investment credit, total sources of fund/ total uses of fund, trade balance/gross domestic
product
[65] Quick ratio, cash ow/sales, cash ow/stock holders equity, cash ow/total assets, cash ow/total borrowings and bond,
change in payable/receivables, change in inventory/current assets, change in inventories turnover, change in payables/
current liabilities, current ratio, current ratio trend, debtos ratio (days) (debtors * 365 days/sales), dividend/capital stock,
nancial expenses/sales, xed assets/(stockholders equity + long-term liabilities), xed assets turnover, xed asset
composition, xed liability ratio, xed ratio, gross prot/sales, growth rate of xed asset, growth rate of net income, growth
rate of ordinary income, growth rate of sales, growth rate of total liabilities, growth rate of total assets, interest coverage
ratio, interest ratio, inventory/current assets, inventories turnover, net income/capital stock, net income/sales, net income/
stockholders equity, net income/total assets, net working capital/total assets, net working capital turnover, operating
income/sales, ordinary income/business capital, ordinary income/sales, ordinary income/stockholders equity, ordinary
income/total assets, payable/current liabilities, payables/inventories, payables/receivables, stock holders equity/total assets,
stock holders equity turnover, total assets turnover, (total borrowings + bonds payable)/total assets, total liability
composition
[66] EBIT/total assets, market capitalization/total debt, retained earnings/total assets, sales/total assets, working capital/total
assets
[67] Average market equity/total capital, auditor, auditor opinion, bond rating, book equity/total capital, quick ratio, cash ow/
xed charges, cash ow/share, cash ow/total debt, cash ow margin, capital expenditure/share, capital lease, cost of goods
sold/sales, current ratio, current liabilities/total liabilities, dividend, earnings/5 years maturity, earnings/total debt, earnings
before interest and taxes (EBIT) drop, EBIT/sales, EBIT/share, EBIT/total assets, EBIT/total tangible assets, xed charge
coverage, interest coverage, inventory turnover, log (interest coverage), log (total assets), long-term debt/equity, margin drop,
market equity/total capital, market value/total liabilities, net available for capital/total capital, net available for total capital/
sales, ROA, net income/total debt, net prot margin, number of employees, operating income/sales, price/earning ratio, quick
assets/sales, receivables turnover, retained earnings/total assets, retained earnings/tangible assets, sales/cash, sales/gross xed
assets, sales/receivables, sales/total assets, sales/total capital, sales/total tangible assets, standard deviation (EBIT/total assets),
standard deviation (log (EBIT/total assets)), total debt/total assets, total debt/total capital, total investment, working capital/
long-term debt, working capital/total assets, worth/total debt
[69] Average salary/employee, borrowing ratio, quick ratio, cash ratio(cash and market securities/total liabilities), cash earnings/
share, cash ow margin, capital employed/employee, capital gearing, coverage debt, creditors ratio (days)(creditor * 365 days/
cost of sales), creditors turnover, current ratio, debtos ratio (days)(debtors * 365 days/sales), debtos turnover, earnings margin,
nancial debt ratio, gross return, income gearing, inventories, net prot margin, net return, operating prot margin, operating
prot/employee, pre-tax prot margin, preferences and loan/equity and reserves, quick assets/total assets, return on capital
employed, return on long-term capital, return on net xed assets, return on shareholders capital, return on shareholders
equity, sales/employee, stock ratio (days), stock turnover, tax ratio, trading prot margin, turnover/assets employed, turnover/
xed assets, turnover/net current assets, working capital/total assets
P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128 7

Table 2 (continued)
Ref # Variables
[70] Auditor opinion, audit qualication, cash/total assets, cash/current liabilities, commercial paper rating, current ratio, current
assets/net sales, current assets/total assets, debt rating, funds ow/total liabilities, net income/total assets, net worth, number of
years of nancial statements in database, number of consecutive years negative net income, number of consecutive years sales
decline, quick assets/current liabilities, sales, sales/total assets, sales/working capital, standard deviation of common stock rate
of return, stock exchange, total liabilities/net worth, total liabilities/total assets, working capital/total assets, yearly dividend
[71] Charge-os/(net operating income + loss provision), commercial and industrial loans/total loans, dividends/net income, equity
capital/adjusted risk assets, gross capital/adjusted risk assets, gross capital/risk assets, gross charge-os/(net operating
income + loss provision), liquid assets/total sources of funds, loans/total assets, loans and leases/total sources of funds, loss
provision/(loans + securities), roa, net income/total assets, net interest margin/earning assets, net interest margin (taxable
equivalent)/earning assets, net liquid assets/total assets, non-interest expenses/operating revenue, operating expenses/operating
revenues, total assets, total operating expenses/operating revenue
[73] Accounts receivables/sales, cash/total assets, current ratio, current assets/total assets, current assets/sales, inventory/cost of
goods sold, long-term debt/total assets, ROA
[74] Cash/current liabilities, investment cash ow/net income, rm size, ROA
[75] Cash/current liabilities, coded to indicate opinion type, current ratio, dividends/net income, rm size, investment cash ow/net
income, leverage, ROA, Operating cash ow/net income, retained earnings/total assets, sales/total assets
[78] Current ratio, current liabilities/total assets, gross prot/total assets, inventories/working capital, (long-term debt + current
liabilities)/total assets, net income/gross prot, net income/net worth, ROA, net worth/net xed assets, net worth/(net worth +
long-term debt), quick assets/current liabilities, working capital/net worth
[79] Break even point ratio, bonds payable, cash ow/interest expenses, cash ow/(previous years short-term loan), cash ow/short-
term loan, cash ow/total debt, cash ow/total loans, capital stock turnover, depreciation ratio, EBIT/sales, xed assets/
(stockholders equity + long-term liabilities), xed assets turnover, xed ratio, gross value added/(property, plant and equip-
ment), gross value added/sales, gross value added/total assets, growth rate of tangible assets, interest coverage ratio, interest
expenses/total borrowings, interest expenses/total expenses, interest expenses/sales, inventories turnover, net income/sales, net
income/stockholders equity, ROA, net interest expenses/sales, operating assets turnover, ordinary income/ordinary expenses,
ordinary income/sales, ordinary income/stockholders equity, ordinary income/total assets, payable turnover, productivity of
capital, solvency ratio, stock holders equity/total assets, stock holders equity turnover, tangible assets turnover, total assets
turnover, (total borrowings + bonds payable)/total assets, variable cost/sales
[83] Current liabilities/current assets, funds provided by operations/total liabilities, log (total assets/GNP price-level index), net
income/total assets, one if total liabilities exceeds total assets, zero otherwise, one if net income was negative for the last two
years, otherwise zero, total liabilities/total assets
[85] Cash ow/total loans, coverage debt, current assets/total assets, current assets-cash/total assets, net income/loans, ROA, net
income/total equity capital, reserves/loans
[86] Quick ratio, nancial expenses/sales, xed assets/(stockholders equity + long-term liabilities), gross value added/tangible xed
assets, gross value added/sales, growth rate of property, plant and equipment, growth rate of sales, growth potential, rm
history, industry position, industry reputation, international competitive advantage, market niche/trend, operating assets
turnover, operating income/total assets, ordinary income/total assets, past payment record, personnel and sta hiring policy,
pricing competitive advantage, prot perspective, quality of management, relationship between labour and capital, size,
stockholders equity/total assets, technology development and quality innovation, total assets turnover, (total
borrowings + bonds payable)/total assets, working conditions and welfare facilities
[88] Current ratio, EBIT/interest expenses, EBIT/total assets, market value of equity/book value of debt, retained earnings/total
assets
[89] Cash at year end/total debt, cash ow/total debt, charge in inventories, charge in net nancials, charge in net other assets and
liability, charge in other current assets, charge in other current liabilities, charge in payables, charge in receivables, current
ratio, current ratio trend, dividend, earnings trend, xed coverage expenditure, long-term debt/net worth, net income/sales,
ROA, net investment ow, net operating ow, quick assets/current liabilities, quick assets/sales, sales trend (number
consecutive years of sales decline, total debt/total assets, trend of cash ow/total debt, trend of net income/sales, trend of net
income/total assets, trend of working capital/sales, working capital/sales
[92] EBIT/total assets, market value of equity/total debt, retained earnings/total assets, sales/total assets, working capital/total assets
[97] Cash/current liabilities, cash/total assets, cash ow/total assets, cash ow/total debt, current ratio, current assets/total assets,
current assets/total sales, current liabilities/equity, EBIT/total assets, equity/sales, inventory/sales, market value of equity/total
capitalization, market value of equity/total debt, ROA, net income/total capitalization, quick assets/current liabilities, quick
assets/total assets, quick assets/total sales, retained earnings/total assets, sales/total assets, total debt/total assets, working
capital/sales, working capital/total assets
[98] GAAP (generally accepted accounting principals) net worth/total assets (GNWTA), repossessed assets/total assets (RATA),
net income/gross income (NIGI), net income/total assets (NITA), cash securities/total assets (CSTA)
[100] EBIT/total assets, market value of equity/total debt, retained earnings/total assets, sales/total assets, working capital/total
assets
(continued on next page)
8 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

Table 2 (continued)
Ref # Variables
[101] EBIT/total assets, market value of equity/total debt, retained earnings/total assets, sales/total assets, working capital/total
assets
[102] Quick ratio, current liabilities/total assets, nancial expenses/sales, liquidity ratio, net income/stockholders equity, operating
income/operating expenses, retained earnings/total assets, stock holders equity/total assets, value added/total cost
[104] Current ratio, EBIT/interest expenses, EBIT/total assets, market value of equity/book value of debt, retained earnings/total
assets
[106] Allowance for loan losses/total loans, asset growth, branch or unit bank, charter, core deposits/total assets, deposit insurance,
earning assets/total assets, federal reserve bank member, gains (losses) from sale of securities/total assets, holding company
aliation, interest income/total assets, (non-interest expenses-salary)/total assets, non-interest income/total assets, non-
performing assets/total assets, o balance sheet commitments/total assets, provision expenses/total loans, regional geographical
region, salary/total assets, total equity/total assets, total interest expenses/total assets, total loans/total assets, total securities/
total assets, volatile liabilities/total liabilities
[108] (Agriculture production and farm loans + real estate loans secured by farm land)/net loans and leases, (cash + US treasury and
government agency obligations)/total assets, capital/assets, commercial and industrial loans/net loans and leases, (federal funds
sold + securities)/total assets, (interest and fees on loans + income from lease nancing)/net loans and leases, loans to
individuals/net loan and leases, net charge os/average loans, ROA, provision for loan loses/average loans, return on average
assets, total expenses/total assets, total income/total expenses, total interest paid on deposits/total deposits, total loans 90 days
or more past due/net loans and leases, total loans and leases/total assets, total loans and leases/total deposits, total non-accrual
loans and leases/net loans and leases
[110] (Agriculture production and farm loans + real estate loans secured by farm land)/net loans and leases, (cash + US treasury and
government agency obligations)/total assets, capital/assets, commercial and industrial loans/net loans and leases, (federal funds
sold + securities)/total assets, (interest and fees on loans + income from lease nancing)/net loans and leases, interest income/
total assets, loans to individuals/net loan and leases, net charge os/average loans, net income/total assets, provision for loan
loses/average loans, real estate loans/(net loan & leases), return on average assets, total expenses/total assets, total interest paid
on deposits/total deposits, total loans 90 days or more past due/net loans and leases, total loans and leases/total assets, total
loans and leases/total deposits, total non-accrual loans and leases/net loans and leases
[114] After tax prot/total assets, cash/total liabilities, working capital/operational expenditure
[115] (Average tangible xed assets-average construction in progress)/number of employees, current ratio, interests, discounts and
bond issue expenses/sales, interests, discount expenses/value added, liquid assets/current liabilities, non-operating expenses/
sales, operating capital/number of employees, operating income/operating capital, ordinary income/sales, value added/
operating capital
[116] Asset (loan) quality, earnings, liquidity, management, miscellaneous
[119] (Cash + US treasury securities + federal funds sold and securities purchased under agreements to resell)/total assets, (certicate
of deposit over $100,000 + federal funds sold and securities purchased under agreements to repurchase)/total assets,
commercial and industrial loans/total loans, doubtful loans/total capital, equity capital/total assets, (nance agriculture
loans + farmers loans + real estate loans secured by farmland)/total assets, loans believed to be uncollectable/total capital,
loans to individuals for household, family and other personal expenditure/total loans, net income/equity capital, ROA, real
estate loans secured by 14 family residential properties/total loans, real estate loans secured by non-farm non-residential
properties/total loans, substandard loans/total capital, total time and savings deposits/total deposits, total assets, total interest
paid on deposits/total deposits, total loans/(equity capital + reserve for loan losses), total loans/savings, total operating
expenses/total assets
[121] EBIT/total assets, market value of equity/total debt, retained earnings/total assets, sales/total assets, working capital/total
assets
[122] Current liabilities/total debts, exploration expenses/total reserves, net cash ow/total assets, total debt/total assets, trend in
total reserves
[124] Debt/gross cash ow, earnings after abnormals /total assets, payout on operating prot before abnormals and tax, pre-tax
prot/total assets, working capital/total assets
[126] Current ratio, EBIT/total assets, market value of equity/total debt, retained earnings/total assets, sales/total assets, working
capital/total assets
[128] Current ratio, ROA, total debt/total assets

the popular neural network architectures. They LVQ has its origin in vector quantization (VQ),
dier in aspects including the type of learning, node the main form of competitive learning neural net-
connection mechanism, the training algorithm, work. In VQ, each of the competitive units corre-
etc. Since, LVQ and Cascor are not very often sponds to a cluster center and the error function is
used networks, their brief overview is presented as the sum of squared euclidean distance between each
follows: training case and the nearest center. LVQ is used to
Table 3
Other details of each paper
Ref # Bank/rm Source of data/country of origin #Samples Techniques used Timeline of dataset
[1] Firms Korea, KIS (Korea Information Service)* 2400 DA, BPNN, RNN1, RNN2, Hybrid 19941997
I, Hybrid II
[2] Banks USA, FDIC Annual report* 100 CNN, SONN, Fuzzy clustering 1991
[4] Firms Italy 3465 BPNN, LDA 19821992 (1 year)
[5] Firms 111 ZETA analysis 19691975

P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128
[6] Firms Spanish 5671 LDA, logit, multi-level perceptron,
Fuzzy rule base
[7] Firms USA 1160 BPNN with novel indicators 3 years
[8] Firms Finnish 74 DA/DA, LA/LA, GA/BPNN, DA/ 19861989
BPNN, LA/BPNN
[9] Firms Korea 662 Auto-associative NN 19942000
[10] Data used by Johnson and Wicherns [50]* LP to quadratic transformation,
QDF
[11] Firms [Standard and Poors COMPUSTAT 237 BPNN, NPDA, logit 19801991
database, Lexis/Nexis, Moodys Industrial
Report, Commerce Clearing Houses Capital
Change Reporting and The Wall Street
Journal, General Council of SEC, CRSP]*
[12] Banks Texas, [FDIC (Federal Deposit Insurance 2067 NN, logit 19851986
Corporation, Shesunho & Co)]*
[13] Firms Data taken from UCI machine learning 37 Fuzzy rough-NN, Fuzzy-NN, crisp-
repository* NN
[14] Firms Data used by Greco et al. [39]* 39 Rough set 1988
[16] Firms Standard and Poors COMPUSTAT 2085 CBR, logit 19751994
database*
[18] Banks Turkish 40 PCA, MDA, logit, probit, IEWS 19942001
(PCA + MDA + logit + Probit)
[19] Banks Belgium, [Chambers of Commerce, CD-Rom 366 MSD (LP + DA), DEA, C5.0 19941996
National Bank of Belgium]*
[24] Firms Standard and Poors COMPUSTAT 150 Loan classication model from OLS 19751976
database* and MDA
[26] Firms Greece, Greek 78 Rough set model using VCR 19801990
[27] Law cases USA 102 BanXupport (IR + CBR)
[29] Firms Data used by Gentry et al. [34] 36 BPNN, logit
[31] Firms Standard and Poors COMPUSTAT 200 RPA, DA 19711981
database*
[36] Firms Data used by Altman [3]* 66 Neuro-fuzzy and rough classiers
[37] Firms Greece, Greek industrial development bank 39 Classical rough set and rough set with 1988
(ETEVA)* dominance relation
[38] Warehouses 12 Rough set approach with order
domains
(continued on next page)

9
Table 3 (continued)

10
Ref # Bank/rm Source of data/country of origin #Samples Techniques used Timeline of dataset
[44] Banks USA, (Federal Financial Institutions 176 Canonical correlation 1987
Examination Council Form 031)*
[45] Firms Data used by Trippi and Turban [111]* 129 Ontogenic NN, DA, Odom & Sharda
method, Back Propagation,
Perceptron, Athena
[46] Several Data used by Liang [68]* FILM, DA, ID3

P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128
[47] Firms Korea 542 DA, BPNN, CBR 19911993
[48] Firms Korea 544 CBFS, DA, BPNN 19911993
[51] Firms Australia, [Aspect Financial Pty Ltds 8241 Mixed logit model, MNL 19962003
Financial Analysis Database, DatAnalysis
Database, Huntleys Delisted Company
Database, ASX market Comparative
Analysis, Australian Securities and
Investment Commission (ASIC)]*
[52] Banks Taiwan 24 DA, BPNN 2000
[53] Firms Standard and Poors COMPUSTAT 50 MDA 19721976
database*
[54] Firms Finnish 1500 Euclidean-SOM, Fisher metric-SOM
[55] Firms Finnish, Kera Ltd* 1137 LDA, LVQ, SOM, RBF-SOM, K-
NN
[57] Banks USA 8977 IEWS (using logit), IEWS (using 19891992, 1 and 2 year
Trait recognition)
[62] Firms Standard and Poors COMPUSTAT 282 Cascor, DA 19701989
database*
[63] Firms Standard and Poors COMPUSTAT 34 FLDF, CBLP, MSD, LPC
database* (combined method)
[64] Firms Standard and Poors COMPUSTAT 364 BPNN 19851995 (1, 2 and 3)
database*
[65] Firms Korea, Korean Stock Exchange* 166 MDA, ID3, MDA-assisted BPNN, 19791992
ID3-assisted BPNN, SOFM-assisted
BPNN
[66] Firms South korea, Korean Investors Service 168 BPNN, DA, SOM 19951998 (2, 3)
(KIS)*
[67] Firms Standard and Poors COMPUSTAT 88 MDA, MDA with gray zone, NN 1984 1988 (3)
database* (perceptron)
[69] Firms UK, London Stock Exchange* 1133 Integration methods of DA, logit, 19801999 (1)
Decision tree and BPNN
[70] Firms [Moodys Bond Record, Standard and Poors 921 Recursive partitioning,
Stock Guide, CRSP tapes and Daily Stock bootstrapping, polytomous probit
Price Record]*
[71] Banks USA, US commercial banks of Federal 5700 Logistic regression 19701976
Reserve System*
[73] Firms USA, Compact Disclosure* 200 Rough set model, Recursive 19861988
partitioning model
[74] Firms USA, Compact Disclosure* 291 Rough set, GA, Rough set + GA 19911997
[75] Firms USA, Compact Disclosure* 291 Rough sets, Auditor 19911997
[78] Firms Greece 110 Fuzzy rule generator, LDA, QDA, 5 and 3
logit, probit
[79] Firms Korea, Koreas largest credit guarantee 1888 SVM, LDA, logit, BPNN 20002002
organization*
[80] Firms [1980 Bankruptcy Yearbook and Almanac, 150 ICPB-MSD, LDA, MSD, ICPB-GA 19871995, 3 years

P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128
Lexis/Nexis Database, Standard & Poors and GA
COMPUSTAT database]*
[83] Firms [Moodys Manual, Standard & Poors 2163 Logit 19701976
COMPUSTAT database, 10-k nancial
statements]*
[85] Banks Spanish 66 BPNN, logit, MARS, C4.5, DA 19971985
[86] Firms South Korea, Industrial bank of Korea* 2144 AHP-CBR, logit-CBR, weighted- 19951998
KNN, Pure-KNN
[88] Firms [Bankruptcy Yearbook and Almanac, Lexis/ 100 GA-based BPNN, BPNN 19891990
Nexis database]*
[89] Banks and 1. Belgian 2. Texas, Data used by Tam and 182 202 48 BPNN with Feature construction 19871989 19851987 (1 and 2) 1980
Firms Kiang [108]* 3. The data used by Rashad (FC), BPNN without FC
et al. [93]*
[92] Firms Moody Industrial Manuals* 129 BPNN, Athena, Perceptron 19751982
[97] Firms Standard and Poors COMPUSTAT* 301 Isotonic separation, DA, LPD, 19962001 (1, 2 and 3 years)
BPNN, Rough sets and OC1
[98] Firms Federal Home Loan Bank board quarterly 862 BPNN, logit 19861987
tapes*
[100] Firms Moodys Industrial Manuals* 129 SOFM, LDA, MLP, Perceptron, 19751985
Athena.
[101] Firms Moodys Industrial Manuals* 129 DA, BPNN, logit 19751982
[102] Firms South Korea 528 GA 19951997
[106] Banks USA 1741 MDA, BPNN, Regulators 1993
[107] Banks Texas 188 BPNN, Factor logistic, DA, K-NN, 19851987 (2 years)
ID3
[108] Banks Texas 202 LDA, logit, K-NN, ID3, feed 19851987
forward NN, BPNN
[112] Firms UK DataStream and FT Excel company 904 Quadratic interval logistic regression 19851994 (1 year)
Research*
[113] Firms Japan 114 BPNN, DA 19701991
[124] Banks USA 3103 GenSoFNN-CRI (S), MCMAC 19802000
[116] Firms Italy 4738 GA, LDA 19821995
[118] Firms German 6667 GANN, GA, BPNN
[119] Banks USA 1900 FA + logit 19801982
[121] Firms Moodys Industrial Manual* 129 BPNN, DA 19751982
(continued on next page)

11
12 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

solve classication problems. Each codebook vector


is assigned to one of the target classes. Each class
may have one or more codebook vectors. A case
is classied by nding the nearest codebook vector
and assigning the case to the class corresponding
to the codebook vector (see comp.ai.neural-nets
FAQ, Part 1 of 7: Introduction) [56]. Fahlman
Timeline of dataset

and Lebiere [28] introduced cascade correlation net-


work, which adds units to the network during train-
1980 1991

ing. Such a network is trained for a while. Then,


19841989

19912001

19721978

without altering the existing weights, one or more


new hidden units are added to the network and then
training resumes. The advantages of Cascor over
backpropagation are: (1) automatic determination
CBR (weighted NN, Pure NN), DA

Simple probit, bivariate probit and

of a network topology resulting in good generaliza-


tion (2) training time reduced by orders of magni-
PNN, PNN without pattern

tude and (3) the ability to rene an existing


normalization, FDA, DA

architecture to take into account new data without


beginning the training process all over again [62].
unweighted probit
Techniques used

BPNN, Logistic

2.1.3. Decision trees


Decision trees form a part of machine learning
an important area of articial intelligence [33,91].
A majority of the decision tree algorithms are used
Note: In the third column * indicates source of data and indicates that the data is not available.

for solving classication problems. However, algo-


rithms like CART (classication and regression
trees) can be used for solving regression problems
#Samples

also. All these algorithms induce a binary tree on


1681
122

44
220

the given training data, which in turn results in a


set of ifthen rules. These rules can be used to solve
the classication or regression problem. A number
American and New York Stock Exchanges*

of algorithms are used for building decision tree


including CHAID (Chi squared automatic interac-
tion detection), CART, Quest and C5.0 [94].
Australia, CD Financial Analysis*
Source of data/country of origin

2.1.4. Rough sets


Rough set theory proposed by Pawlak [87] is
based on the assumption that with any object of
the given universe there is some information associ-
ated and objects characterized by similar informa-
tion are indistinguishable or indiscernible. The
indiscernibility relation indicates that we are unable
USA

USA

to deal with single objects but we have to consider


clusters of indiscernible objects or equivalence clas-
ses of the indiscernibility relation. In rough set the-
ory, a pair of precise concepts viz., lower and the
Bank/rm

upper approximations replaces any vague concept.


Table 3 (continued)

Firms

Firms
Firms
Firms

Approximations are two basic operations in the


rough set theory. Rough sets can be applied for
inducing decision rules from data, to solve classi-
Ref #

cation problems. The induced decision rules are cat-


[122]

[123]
[126]
[128]

egorized into exact and possible.


P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128 13

2.1.5. Case-based reasoning practices. Each DMUs eciency is then measured


Case-based reasoning (CBR), a branch of arti- relative to this frontier. In other words, DEA
cial intelligence, is intuitively similar to the cognitive assesses the eciency of each DMU relative to all
process humans follow in problem solving [58]. the DMUs in the sample, including itself. This rela-
When people confront a new problem, they often tive eciency is calculated by obtaining the ratio of
depend on past similar experiences and reuse or the weighted sum of all outputs and the weighted
modify solutions of these experiences to generate a sum of all inputs. The weights are selected so as to
possible answer for the problem at hand. The hall- achieve Pareto optimality for each DMU. An
mark of CBR is its capability to give an explanation appealing aspect of DEA is that it does not require
for its decision based on previous cases. Citing rele- price or cost data. Also, the DEA is invariant to
vant previous experiences or cases is a way to justify scaling of variables. The DEA helps to identify inef-
a position [58] in human decision-making. Compre- cient DMUs as well and amounts of ineciency of
hensibility of the decision is often crucial in solving inputs and/or outputs.
nancial problems. When a company is identied as
failing, CBR can give examples of similar compa-
nies that failed in the past as a justication for its 2.1.8. Soft computing
prediction. The heart of the CBR is the nearest The paradigm of soft computing or computational
neighbour algorithm [96]. intelligence refers to the seamless integration of dif-
ferent, seemingly unrelated, intelligent technologies
2.1.6. Support vector machines such as fuzzy logic, neural networks, genetic algo-
Support vector machines (SVM) introduced by rithms, machine learning (case-based reasoning
Vapnik [115] use a linear model to implement non- and decision trees subsumed), rough set theory
linear class boundaries by mapping input vectors and probabilistic reasoning in various permutations
nonlinearly into a high-dimensional feature space. and combinations to exploit their strengths. This
In the new space, an optimal separating hyperplane term was coined by Zadeh [125] in the early 1990s
(OSH) is constructed. The training examples that to distinguish these technologies from the conven-
are closest to the maximum margin hyperplane are tional hard computing that is inspired by the
called support vectors. All other training examples mathematical methodologies of the physical sci-
are irrelevant for dening the binary class bound- ences and focused upon precision, certainty and
aries [22,115]. SVM is simple enough to be analyzed rigor, leaving little room for modeling error, judg-
mathematically. In this sense, SVM may serve as a ment, ambiguity, or compromise. In contrast, soft
promising alternative combining the strengths of computing is driven by the idea that the gains
conventional statistical methods that are more achieved by precision and certainty are frequently
theory-driven and easy to analyze and machine not justied by their costs, whereas the inexact com-
learning methods that are more data-driven, distri- putation, heuristic reasoning and subjective decision
bution-free and robust. Recently, the SVM has been making performed by human minds are adequate
used to nancial applications such as credit rating, and sometimes superior for practical purposes in
time series prediction and insurance claim fraud many contexts. Soft computing views the human
detection. These studies reported that SVM was mind as a role model and builds upon a mathemat-
comparable to and even outperformed other classi- ical formalization of the cognitive processes those
ers including BPNN, CBR, MDA and logit in humans take for granted [125]. Within the soft com-
terms of generalization performance. puting paradigm, the predominant reason for the
hybridization of intelligent technologies is that they
2.1.7. Data envelopment analysis are found to be complementary rather than compet-
Data envelopment analysis (DEA) is a non-para- itive in several aspects such as eciency, fault and
metric performance assessment methodology intro- imprecision tolerance and learning from example
duced by Cooper et al. [21] to measure the relative [125]. Further, the resulting hybrid architectures
eciencies of organizational or decision-making tend to minimize the disadvantages of the individual
units (DMUs). The DEA applies linear program- technologies while maximizing their advantages.
ming to observing inputs consumed and outputs Some of the soft computing architectures employed
produced by decision-making units and con- are neurofuzzy, fuzzyneural, neurogenetic,
structs an ecient production frontier based on best geneticfuzzy, neurofuzzygenetic, roughneuro,
14 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

etc. Multi-classication systems or ensemble classi- was taken from Call & Income reports and exami-
ers are also treated as soft computing systems. nation reports for 1900 commercial banks in some
states of USA. He demonstrated that his combined
3. Statistical techniques method of factor analysis and logit estimation was
promising in evaluating banks condition. Karels
Altman et al. [5] developed a new bankruptcy and Prakash [53] conducted a study in a threefold
classication model called Zeta analysis, which manner; (i) Investigated the normality condition of
incorporated comprehensive inputs. The data sam- nancial ratios required by the MDA technique;
ple consists of 111 rms with seven variables each. (ii) when these ratios are non-normal, they con-
They found that the ZETA model outperformed structed ratios which are either multivariate normal
alternative methods in terms of expected cost crite- or almost normal; (iii) using these ratios to compare
ria. Classication accuracy of this model ranged the prediction results of DA with that of other stud-
from 96% for one period prior to bankruptcy to ies. They used a random sample of 50 companies
70% for ve periods prior to bankruptcy. Martin obtained from the COMPUSTAT data tapes. They
[71] presented logistic regression to predict proba- reported 96% classication rate for non-bankrupt
bility of failure of banks based on the data rms and 54.5% for bankrupt rms.
obtained from the Federal Reserve System, data Haslem et al. [44] determined the implication of
sample. Ohlson [83] employed the logit model to the foreign and domestic strategies reected in the
predict rm failure. The data was obtained from balance sheets of US commercial banks. Also, they
Moodys Manual, Compustat data tapes and 10- determined the impact of strategies on the prot-
K nancial statements. The classication accuracy ability performance. For the analysis, they used
reported by him was 96.12%, 95.55% and 92.84% the 1987 balance sheet data from the call reports
for prediction within one year, two years and one of 176 large US banks having both foreign and
or two years respectively. Dietrich and Kaplan domestic oces. They used canonical correlation
[24] developed a simple three-variable linear model analysis to analyze balance sheet for the large and
to classify loan risks. He compared his model with very large bank samples and generated canonical
(i) Altman model and (ii) Wilcox bankruptcy pre- variate scores. Kolari et al. [57] developed early
diction [120] model and got better performance warning system (EWS) based on logit analysis and
over them. They used six independent variables Trait recognition for large US banks. The logit
and one dependent variable suggested by individual model correctly classied over 96% of the banks 1
experts. The model gave better accuracy than previ- year prior to failure and 95% of the banks 2 year
ous bankruptcy models in predicting the riskiness prior to failure. For developing trait recognition
of loan. model, they used half of the original sample.
Zmijewski [128] examined two estimation biases Because of this reduction in the sample size it has
for nancial distress models on non-random disadvantages over logit model. In both 1 year
samples. The rst bias results from oversampling and 2 year prior to failure data classication accu-
distressed rms and falls under the topic of choice- racy of trait model was 100%. It was concluded that
based sample biases. The second results from using trait recognition outperformed logit model in terms
a complete data sample selection criterion and of type-I and type-II errors.
falls under the topic of sample selection biases. Recently, Jones and Hensher [51] presented
The data used in the study was obtained from the mixed logit model for rm distress prediction and
American and New York Stock Exchanges. The compared it with multinomial logit models
data set consisted of estimation sample of 40 bank- (MNL). They modeled nancial distress problem
rupt and 800 non-bankrupt rms and a prediction in three states viz., state 0: non-failed rms; state
sample of 41 bankrupt and 800 non-bankrupt rms. 1: insolvent rms, state 2: rms which led for bank-
The choice-based sample was examined using ruptcy. They developed two samples for model esti-
unweighted probit and weighted exogenous sample mation and validation. The results are too detailed
maximum likelihood. The sample selection bias to present here. They concluded that mixed logit
issue was examined using simple probit and bivari- obtained substantially better predictive accuracy
ate probit assessment. West [119] used the factor than multinomial logit models. Canbas et al. [18]
analysis to create composite variables to describe proposed an integrated early warning system
banks nancial and operating characteristics. Data (IEWS) by combining DA, logistic regression, pro-
P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128 15

bit and principal component analysis (PCA), that performance with a logistic model. They considered
helped in examination and detection of banks with 29 variables from the CAMEL categories and per-
serious problems. PCA explored three nancial formed stepwise regression that identied ve vari-
components, which signicantly explained the ables. Results are too detailed to present here.
changes in the nancial condition of banks. With They concluded that BPNN outperformed logistic
these three factors they employed DA, logistic regression. Sharda and Wilson [101] compared the
regression and probit models. Then, by combining BPNN with MDA based on resampling technique.
all these together they constructed an IEWS. They They used Altmans ve variables. They used
used the data of Turkish banks available in website BPNN with ve input nodes, 10 hidden nodes and
(http://www.tbb.org.tr/english/bulten/yillik/2000/ two output nodes. They concluded that BPNN out-
ratios.xls). The results are too elaborate to present performed DA in all cases. Fletcher and Goss [29]
here. They concluded that IEWS has more predic- used the BPNN for predicting bankruptcy of rms
tive ability than other models. and compared it with logistic regression. They
employed v-fold cross-validation technique. The
4. Neural networks data of 36 companies was drawn from Gentry
et al. [34]. They used three independent variables.
This section reviews the papers where dierent BPNN prediction performance was 82.4% whereas
topologies of neural networks (NN) are proposed logistic regression produced only 77%. Altman
and compared with other techniques. This section et al. [4] compared the performance of LDA with
is split into three sub sections covering the applica- BPNN in distress classication. They used 10 nan-
tions of (i) back propagation trained NN (BPNN), cial ratios. The data sample consisted of three types
(ii) self-organizing feature map (SOM) and (iii) of rms viz., (i) healthy, (ii) unsound and (iii) vul-
other NN topologies such as probabilistic NN, auto nerable. Three sample data sets were used for this
associative NN and cascade correlation NN. study. They obtained best results with BPNN hav-
ing three layers: initial hidden layer of 10 neurons,
4.1. Back propagation trained neural network second hidden layer with four neurons and an out-
(BPNN) put layer of one neuron. They concluded that
LDA marginally outperformed BPNN.
Tam [107] employed BPNN for bankruptcy pre- Wilson and Sharda [121] compared the predictive
diction. Data were obtained from Texas banks, one accuracy of BPNN with that of DA. They used Alt-
year and two years prior to failure. He selected the man [3] variables. They concluded that BPNN out-
variables based on CAMEL criteria of FDIC. He performed other methods over all test samples.
showed that BPNN oered better predictive accu- Tsukuda and Baba [113] performed the prediction
racy than other methods viz., DA, factor-logistic, of bankruptcy using BPNN with one hidden layer
K-nearest neighbour (K-NN) and ID3 [77]. Tam and concluded that BPNN outperformed the DA.
and Kiang [108] compared the performance of (i) Leshno and Spector [67] compared various NN
LDA, (ii) logistic regression, (iii) K-NN, (iv) ID3, models and DA and explained their prediction
(v) feedforward NN (net0) and (vi) BPNN (net10) capabilities in terms of data span, learning tech-
on bankruptcy prediction problems. BPNN outper- nique and number of iterations. This study used
formed other techniques for one-year prior training the perceptron like network for input features of
sample, whereas for two-year prior training sample two types viz., (i) functional expansion (FE), (ii)
DA outperformed others. However, BPNN outper- joint activation (JA) and investigated their eect
formed others in both the one-year prior and the on generalization. They used 41 nancial ratios
two-year prior holdout samples. In jackknife including Altman [3] variables. They concluded that
method [61] also, BPNN outperformed others in NN outperformed the Z-score model. Rahimian
both the one-year prior and the two-year prior hold- et al. [92] compared the performance of (i) BPNN,
out samples. So, with the evidence of jackknife (ii) Athena, an entropy-based neural network and
method, they concluded that NN outperformed (iii) single layer perceptron on the bankruptcy pre-
the DA methods. diction problem. They compared them with Odom
Salchenberger et al. [98] presented a BPNN and Shardas [82] BPNN and discriminant analysis
to predict the probability of failure for savings also. The accuracies obtained on the test data were
and loan associations (S&Ls). They compared its (i) discriminant analysis produced 74.54%, (ii)
16 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

Odom and Shardas test produced 81.81%, (iii) respectively. However, BPNN model correctly clas-
Athena yeilded 81.81%, (iv) perceptron yielded sied 81.4% and 78.25% of regional and community
81.81% and (v) BPNN produced 81.81%. The banks. They concluded that BPNN outperformed
BPNN implemented in this study outperformed other two models in identifying under performance
Odom and Shardas [82] BPNN in terms of speed. banks. Lam [64] integrated fundamental and techni-
Barniv et al. [11] compared BPNN, multi-state cal analysis in BPNN for nancial performance pre-
ordered logit and non-parametric multiple discrimi- diction. The predictors included 16 nancial and 11
nant analysis (NPDA). They developed a model for macroeconomic variables. She concluded that
classifying rms into three states viz. acquired, BPNN signicantly outperformed the minimum
emerging and liquidated. They designed two models benchmark based on highly diversied investment
viz., (i) twelve variable model and (ii) ve variable strategy. Also, incorporation of previous years
model. These two models outperformed Ohlsons nancial data in the input vector for BPNN could
[83] nine variable model. They concluded that signicantly increase the return level, thereby, dem-
BPNN outperformed NPDA and logit models. Bell onstrating the benets of integrating fundamental
[12] compared logistic regression and BPNN in pre- analysis with technical analysis via BPNN. She also
dicting bank failures. In this study, he used 28 can- extracted rules from trained neural network and
didate predictor variables. The architecture of found that they outperformed the neural networks
BPNN was 12 input nodes, six hidden nodes and per se. Further, they performed as accurately as
one output node. He concluded that neither the the maximum benchmark. Lee et al. [66] compared
logit nor the BPNN model dominated the other in BPNN with self-organizing feature map (SOM),
terms of predictive ability. But, for complex decision DA and logistic regressions. The data sample con-
processes BPNN was found to be better. sisted of 168 Korean rms taken from the Security
Piramuthu et al. [89] designed a method called and Exchange Commission (SEC) lings stored in
feature construction (FC) and used it with BPNN an on-line database of the Korea Investors Service
for bankruptcy prediction. They used the Belgian (KIS) Inc. (www.kisrating.com). The fourfold
Bankruptcy data of 182 banks, Tam and Kiang cross-validation testing was used for all the models.
[108] data and Rashid and EI-Sheshai [93] data of The results are too detailed to present here. They
48 banks. They concluded that BPNN with FC out- concluded that the BPNN outperformed the all
performed the plain BPNN in all datasets. Zhang other techniques.
et al. [126] used generalized reducing gradient
(GRG2) trained three-layered NN for bankruptcy 4.2. Self-organizing maps (SOM)
prediction. They used v-fold cross-validation tech-
nique for testing. They reported that overall classi- Lee et al. [65] proposed three hybrid BPNN viz.,
cation rates of GRG2 trained NN ranged from (i) MDA-assisted BPNN (ii) ID3-assisted BPNN
77.27% to 84.09% whereas that for logistic regres- and (iii) SOM-assisted BPNN for predicting bank-
sion ranged from 75% to 81.82%. They concluded ruptcy in rms. The data of 166 rms is taken from
that GRG2 trained NN outperformed logistic the Korea Stock exchange. They selected 57 nan-
regression. Atiya [7] reviewed the applications of cial variables. They concluded that hybrid neural
NN in bankruptcy prediction and developed an network models performed better than MDA and
NN. He developed novel indicators for the NN, tak- ID3. Serrano-Cinca [100] compared the perfor-
ing cue from Merton [76]. For the study he collected mance of SOM with LDA and BPNN in nancial
data from defaulted and solvent US rms. He diagnosis. The data set consisted of Altmans [3]
reported a prediction accuracy of 84.52% for the variables. The architecture of SOM used was 4 input
in-sample set and 81.46% for the out-of-sample nodes and 144 output nodes arranged in a 12 12
set. He showed that the use of the indicators in addi- square grid in order to accommodate 74 patterns
tion to nancial ratios provided signicant in data sample. He proposed two hybrid neural sys-
improvement. tems viz., (i) a combination of LDA with SOM,
Swicegood and Clark [106] compared DA, where LDA calculated the Z-score for each rm,
BPNN and human judgment in predicting bank fail- which was superimposed onto SOM to obtain iso-
ures. The variables were taken from the bank call solvent regions, (ii) a combination of BPNN with
reports. Overall, MDA model correctly classied SOM. The LDA obtained 74.5%, while other mod-
86.4% and 79.5% of regional and community banks els such as Rahimian et al.s MLP [92], Odom and
P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128 17

Shardas MLP [82], Perceptron model, Athena discriminant analysis (FDA) to solve bankruptcy
Model and his own MLP obtained 81.8%. The pro- prediction problem. They compared the original
posed system outscored Z-score analysis in provid- PNN and PNN without pattern normalization
ing intuitive visual graphics that give information (PNN*) and FDA with DA and BPNN. The Data
on the risk of bankruptcy, the nancial characteris- used was taken from Platt et al. [90]. The rst four
tics of rm and that type of rm it is similar to. ratios were deated to remove the dierences in
Kiviluoto [55] used SOM and proposed its variants ratio values over time caused by uctuations in
for rm bankruptcy prediction. He compared three related factors. The results are too detailed to pres-
dierent SOM-based classiers viz., SOM-1, SOM-2 ent here. They concluded that the PNN* and BPNN
and RBF-SOM hybrid with LDA, learning vector with non-deated data achieved better classication
quantization (LVQ) and K-NN. The parameters rates. FDA produced better classication results
for each classier were determined using v-fold with deated data. They found that deation
cross-validation technique. He modied the LVQ improved the discrimination ability of some of the
algorithm to accommodate the NeymanPearson prediction models as observed in [90]. Baek and
classication criteria. This NeymanPearson LVQ Cho [9] proposed the auto-associative neural net-
not only speeded up the convergence but also work (AANN) for Korean rm bankruptcy predic-
increased the classication accuracy. He used the tion. They trained the AANN with only solvent
data segment from Kera Ltds customer companies rms data. Then they applied the test data contain-
for the study. He concluded that NeymanPearson ing both solvent and insolvent rms. So, any solvent
LVQ is more useful than one minimizing the total rm data that shared common characteristics with
number of misclassications. He concluded that the training data resulted in small error at the out-
other classiers outperformed SOM-1. He also con- put layer while the bankrupt rms data resulted in
cluded that RBF-SOM performed slightly better a large error at the output layer. AANN yielded
than other classiers. Kaski et al. [54] introduced classication rates of 80.45% for solvent and
Fisher information matrix based metric and imple- 50.6% for defaulted rms. However, the 2-class
mented SOM with it. They used the new method BPNN produced classication rates of 79.26% for
to understand the non-linear dependencies between solvent and 24.1% for defaulted rms. Therefore,
bankruptcies and nancial indicators. The depen- they concluded that AANN outperformed 2-class
dencies were converted into a metric of the input BPNN.
space and the SOM was used to visualize the
dependencies in a concise form. They obtained 23 5. Case-based reasoning techniques
nancial indicators from Finnish small and med-
ium-sized enterprises. They computed the accuracy Bryant [16] designed a CBR system for bank-
of SOMs in the Euclidean metric (SOM-E) and in ruptcy prediction. He compared it with Ohlsons
the Fisher metric (SOM-F) in representing the prob- [83] logit model. CBR cluster trees were created with
ability of bankruptcy, measured by the likelihood of three case libraries viz., model-I, model-II and
data at the location of the best matching SOM model-III. They consisted of 25 nancial variables
units. He concluded that the SOM-F performed bet- for one year, two year and three years data. The
ter than the SOM-E. results reported by him are too elaborate to mention
here. He concluded that logit outperformed CBR in
4.3. Other neural network topologies terms of less Type-I accuracy. Jo et al. [47] used
MDA, CBR and BPNN to predict bankruptcy.
Lacher et al. [62] proposed a cascade-correlation They considered the data taken from Korean rms.
neural network (Cascor) for classifying nancial Variables were selected using the dimension reduc-
health of a rm. Altmans ve nancial ratios were tion techniques such as stepwise selection and t-test.
used. Data was collected from Standard and Poors The average hit ratio of DA, BPNN and CBR was
COMPUSTAT nancial database. He compared 82.22%, 83.79% and 81.52%, respectively. They con-
the performance of the Cascor model with that of cluded that BPNN outperformed DA and CBR and
Altman Z-score model. They concluded that the DA outperformed CBR. Park and Han [86] pro-
Cascor model consistently yielded higher overall posed analytic reasoning model called the K-NN
classication rates. Yang et al. [122] proposed with analytic hierarchy process (AHP) feature
PNN without pattern normalization and Fisher weight approach for bankruptcy prediction. They
18 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

proposed CBR for indexing and retrieving similar 7. Evolutionary approaches


cases. The AHP-weighted K-NN was compared
with pure K-NN algorithm. For this study, they Varetto [116] employed a genetic algorithm (GA)
used the data obtained from Industrial Bank of for bankruptcy prediction and compared its perfor-
Korea. They used both nancial and non-nancial mance with that of LDA. He conducted the study
variables. The classication accuracy of pure K- for: (i) one year prior to bankruptcy data and (ii)
NN approach was 68.3% whereas the hybrids three years prior to bankruptcy data. He reported
Logit-CBR and AHP-K-NN-CBR produced that for case (i), the genetic linear function yielded
79.2% and 83.0%, respectively. They concluded that 92% classication rate for bankrupt companies and
weighted K-NN hybrid model outperformed other LDA yielded 90.1%. However, in case (ii), he
models. observed that LDA outperformed the genetic linear
Yip [123] used CBR with K-NN to predict Aus- function in the case of sound companies. He also
tralian rm business failure. She used the statistical inferred that the LDA has a higher stability and gen-
evaluations for assigning the relevancy of attributes eralization power. Nanda and Pendharkar [80]
in the retrieval phase of algorithm. She compared incorporated misclassication cost matrix into an
the performance of CBR + K-NN with that of evolutionary classication system. Using simulated
DA. The overall accuracy of CBR with weighted and real-life bankruptcy data, they compared the
K-NN, CBR with pure K-NN and DA were proposed method with LDA, a goal programming
90.9%, 79.5% and 86.4%, respectively. She con- and a GA-based classication without the asymmet-
cluded that CBR with weighted K-NN was better ric misclassication costs. For bankruptcy training
than DA. set the classication accuracy of integrated cost pref-
erence based mininized sum of deviations (ICPB-
6. Decision trees MSD) and integrated cost preference based GA
(ICPB-GA) outperformed LDA, MSD and GA.
In this section, we review the works reported on For simulated holdout set the ICPB-GA outper-
the application of decision trees. Decision trees use formed others. They concluded that the ICPB-
recursive partitioning algorithm to induce rules on MSD or ICPB-GA might be promising when
a given data set. Marais et al. [70] proposed recur- compared to traditional MSD or GA. Shin and
sive partitioning algorithm (RPA) for predicting Lee [102] also proposed a GA-based approach for
bankruptcy in rms. They used (i) recursive parti- bankruptcy prediction. The rules generated by GA
tioning technique and (ii) bootstrapping. They were easily understood and could be used as expert
applied polytomous probit and recursive partition- systems. Data used contained 528 rms. The ve
ing to the data sample. The results reported are rules generated by GA got 80.8% accuracy. They
too detailed results to be presented here. However, concluded that GA could successfully learn linear
it can be inferred that when all the variables were relationship among input variables.
used, polytomous probit outperformed recursive
partitioning in terms of expected misclassication 8. Operational research
rates in both resubstitution and bootstrap methods.
Frydman et al. [31] presented the application of Banks and Prakash [10] proposed the linear pro-
RPA to bankruptcy prediction and compared it gramming heuristic to a quadratic transformation
with the DA and the analysis was carried out for of data to predict rm bankruptcy prediction. Qua-
misclassication cost of C12 ranging from 1 to 70, dratic discriminant function (QDF) got three mis-
where C12 denotes the cost of misclassifying sample classications whereas Johnson and Wichern [50]
belonging to group 1 to group 2. They constructed got 46. They concluded that the quadratic transfor-
two variants of discriminant functions viz., DA1 mation method outperformed the QDF. Lam and
and DA2 and compared them with two RPA mod- Moy [63] compared dierent DA methods and pro-
els viz., (i) RPA1 with relatively complex tree and posed the their combination to predict the classica-
(ii) RPA2 with smallest v-fold cross-validation risk. tion of new observations. For testing this hybrid
RPA1 model outperformed DA1 and DA2 models technique, they used simulation experiment. Dis-
for all costs. Also, RPA2 tree turned out to be sub criminant analyses taken for this study were FLDF
tree of RPA1 tree for every cost. They showed that (Fisher linear discriminant analysis), cluster-based
RPA2 had larger resubstitution risk. LP (CBLP) and MSD (minimized sum of devia-
P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128 19

tions). Their simulation generated data from con- the smaller reduct of {A1, A4} obtained by the pro-
taminated multivariate normal distribution for the posed approach as against the reduct {A1, A2, A3
rst 24 cases and for remaining 10 cases contained and A4} obtained by original rough set analysis
non-contaminated data to test the robustness of obtained. The decision rules obtained from the pro-
LPC, which combined classication results of dier- posed approach gave more synthetic representation
ent DA. Their combined method outperformed of knowledge contained in the information.
other DA methods. LPC was reformulated as a Greco et al. [37] presented a new rough set
mixed-integer programming model, which mini- method based on approximation of a given partition
mized weighted number of misclassications. Cielen of set of rms into pre-dened and ordered catego-
et al. [19] compared the performance of minimized ries of risk by means of dominance relation in place
sum of deviations (MSD), data envelopment analy- of indiscernibility relation for the evaluation of
sis (DEA) model and a rule induction (C5.0) model bankruptcy in rms. Dominance relation was
on bankruptcy prediction problem. Here MSD was proposed in place of indiscernibility relation in
a combination of linear programming (LP) and DA. [3943,103]. The classical approach based on indis-
The dataset was taken from National Bank of cernibility relation is applied to classication prob-
Belgium. The MSD, DEA and C5.0, obtained clas- lems with regular attributes and preferentially
sication rates of 78.9%, 86.4% and 85.5% respec- non-ordered decision classes, whereas, the domi-
tively. They concluded that DEA outperformed nance-based rough set approach handles prefer-
C5.0 and MSD model. Kao and Liu [52] formulated ence-ordered domains of attributes (criteria) and
a DEA model for interval data for evaluating the preference-ordered decision classes. The domi-
performance of banks. This study made advance nance-based rough set approach performs very well
predictions of the performance of 24 Taiwan banks on nancial data. It is also the only data mining
based on uncertain nancial data presented in approach handling preference order in data. The
ranges. They presented the prediction of eciency data set was obtained from Greek industrial devel-
scores also in ranges. Among 24 banks, two banks opment bank (ETEVA). The data sample was clas-
got the smallest predictive eciency scores of sied into (i) unacceptable, (ii) uncertainty and (iii)
0.7358 and 0.7584 because these two banks suered acceptable. They compared classical rough set with
from Asian nancial crisis and had many bad debts. their proposed method. They showed that the pro-
They showed that DEA predicted the bank perfor- posed method gave smaller number of reducts than
mance based on their nancial forecasts. classical rough set approach based on indiscernibil-
ity. The quality of the approximation obtained by
9. Rough sets classical rough set approach and the proposed
method based on dominance were 1 and 0.949
Greco et al. [38] proposed a new rough set respectively. Also, the decision rules obtained from
approach for solving bankruptcy prediction prob- the approximation by dominance relation gave a
lem considering the criteria of attributes with more synthetic representation of information con-
ordered domains. This approach was similar to tained in decision tables. They showed that rules
the original rough set analysis in that it used the based on dominance relation are better to sort
approximation of partitioning the objects in some new actions than the rules based on indescernibility
pre-dened category. However, it employed both relation. Dimitras et al. [26] also used rough set the-
dominance relation and indiscernibility relation. ory for predicting business failures. They used VCR
They represented warehouse making loss with class (valued closeness relation), which prevents major
CL1 and warehouse making prot with class CL2. dierences on one attribute from being compen-
The classical rough set approach got the accuracy sated by number of minor dierences on other attri-
of approximation of 0.6 and 0.7 for CL1 and CL2 butes. They used the data collected from large
respectively and the quality of classication was number of Greek rms. VCR nearest rules correctly
0.83. However, the proposed approach got the accu- classied 60% of banks not classied by exactly
racy of approximation of 0.33 and 0.6 for CL1 and matching rules. They compared the accuracy of
CL2 respectively and the quality of classication rough set based approach with DA and logit and
was 0.67. They concluded that the proposed concluded that the rough set approach outper-
approach showed improvements over the original formed the other two in revealing the relevant attri-
rough set analysis. This improvement be seen by butes in evaluating the rm failure risk. McKee [73]
20 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

developed a rough set based bankruptcy prediction competitive neural network and (ii) SOM. The
model. Rough set analysis produced better results results are too detailed to be reported here. They
when the attribute domains for continuous variables concluded that both fuzzy clustering and SOM are
were nite sets of low cardinality. Here, the vari- good tools in identifying potentially failing banks.
ables identied by recursive partitioning method Andres et al. [6] proposed fuzzy rule based classiers
were used to develop rough set based model which for bankruptcy prediction problem. They compared
yielded 88% accuracy in predicting bankruptcy on the performance of LDA and logistic regression
a 100-company holdout sample. This was superior with multi-layer perceptron and fuzzy-rule based
to the original recursive partitioning model, which classiers. They also used Monte Carlo simulation
was only 65% accurate on the same data set. to measure the eects of sample size variations on
Bioch and Popova [14] proposed a modication the performance of classiers. The distinctive fea-
of the rough set approach for bankruptcy predic- tures of this study were: (i) for each classier and
tion. They used monotone extensions, decision lists a wide range of sample sizes, average error rates
and dualizations to compute classication rules that were estimated from the results of a large number
cover the whole space. The dataset used here were of Monte Carlo simulations. (ii) The focus was on
taken from [105]. They computed the monotone business protability analysis, which was not con-
decision trees for the dataset. They concluded that sidered earlier. (iii) Their classication problem
there was close relationship between the decision had a low separability degree. (iv) A slight variant
rules obtained using the rough set approach and of the class of additive fuzzy systems with Gaussian
the prime implicants of the maximal extension. membership functions and consequent normalized
McKee [75] compared rough set prediction capabil- to be probabilities was tested. They used the data-
ity with actual auditor signaling rates for US com- base of commercial and industrial rms located in
panies. He developed two dierent models with Spain. The results are too elaborate to present here.
four variables from the 11 possible variables. These They concluded that MLP and fuzzy rule based
two models achieved classication accuracy of 61% classier outperformed LDA and logistic regression.
and 68% on the validation set. Auditors achieved Min and Lee [79] proposed SVM [115] for bank-
classication accuracy of 66%. The model accuracy ruptcy prediction. He proposed grid-search tech-
rate in this study was signicantly lower than the nique using vefold cross validation to nd out
one reported in previous studies employing both the optimal parameter values of kernel function of
rough sets theory and other methods. This was SVM. He compared the SVM with MDA, Logit
because the samples employed here were more real- and BPNN. He used two kernels for SVM viz., (i)
istic than prior studies. RBF kernel and (ii) polynomial kernel. The classi-
cation rates of MDA, logit, BPNN and SVM for
10. Other techniques training data were 78.80%, 79.86%, 85.24% and
88.01%, respectively and for holdout data 79.13%,
In this section, the papers applying fuzzy tech- 78.30%, 82.53% and 83.06%, respectively. They con-
niques, SVM and isotonic separation are reviewed. cluded that the SVM outperformed BPNN, MDA
Since these techniques found relatively fewer appli- and logit models for the training and holdout data.
cations in bankruptcy prediction, all of them are Ryu and Yue [97] introduced isotonic separation
grouped under this category. Michael et al. [78] pro- for prediction of rm bankruptcy. They compared
posed fuzzy rule generator method for bankruptcy isotonic separation with (i) DA, (ii) linear program-
prediction and compared it with LDA, QDA, logit ming discrimination (LPD), (iii) BPNN, (iv) LVQ,
analysis and probit analysis. They used two samples (v) rough set analysis and (vi) oblique decision tree
of data from Greek rms. Fuzzy rule based classier method (OC1). They used three feature selection
got 7.48% of type-I error, 44.83% of type-II error methods viz., (i) backward sequential elimination
and overall error was 26.16%. The overall classica- (BSE) method with isotonic separation and LP,
tion error of LDA, logit and probit analysis was (ii) stepwise discriminant analysis (SDA) method
29.35%, 27.15% and 30.63%, respectively. They con- with DA and (iii) mutual information-based
cluded that fuzzy rule based classier outperformed (MIB) method with BPNN. They employed 10-fold
other methods. Alam et al. [2] proposed fuzzy clus- cross validation for testing their method. They con-
tering for identifying potentially failed banks and cluded that the rough set method was the top per-
then compared it with two SOM networks viz., (i) former followed by OC1 and LVQ. BPNN,
P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128 21

logistic discrimination and probit methods per- (GANN), where dierent parameters of NN such
formed worse than others. Results are too detailed as topology, connection weights and input variables
to present here. They concluded that isotonic sepa- selection were encoded. They used the data of Ger-
ration outperfomed other methods for short-term man corporations. The performance of GANN was
bankruptcy prediction. compared on the basis of the Type-II error, while
the costlier Type-I error was kept constant. Using
11. Soft computing techniques all input variables, BPNN achieved a Type-II error
of 44.2% on the test sample. The genetic algorithm
In this section, papers employing various hybrid reduced the Type-II error from 42.6% to 36.1%.
intelligent techniques to solve bankruptcy predic- Optimized GANN achieved Type-II error of
tion problem are reviewed. These papers developed 46.6% on validation set due to over selection
new hybrid systems combining the intelligent tech- eect. Optimized neural network did not outper-
niques and sometimes some proven statistical mod- form earlier solutions on the validation set.
els. Three varieties of soft computing architectures Jo and Han [48] designed a new architecture by
have been employed so far: (i) ensemble classier hybridizing case-based forecasting (CBFS), BPNN
[96], where techniques such as BPNN, DA, logistic and DA, which achieved higher prediction accuracy
regression, MARS, C4.5, fuzzy logic, rough set than these individual models operated in stand-
based approach, GA, GP were employed to solve alone mode. Three types of data viz., training,
the problem in a stand-alone mode and then their testing and generalization, collected from Korean
results are combined through an arbitrator which companies, were used. The proposed hybrid intelli-
performs simple majority voting or weighted major- gent system comprised a linear combination of the
ity voting schemes or a linear combination of pre- following ve models viz., (i) DA, (ii) BPNN1
dictions; (ii) an intelligent technique is used for (which is BPNN with one hidden layer), (iii)
feature selection task and another intelligent tech- BPNN2 (which is BPNN with two hidden layers),
nique performs classication by taking the selected (iv) CBFS1 (which uses a similarity measure to
features and (iii) tightly integrated hybrid systems determine the number of base cases) and (v) CBFS2
such as GA trained NN, neuro-fuzzy, GA-neuro- (which uses all the base cases). They performed
fuzzy, etc. The papers reviewed in this section numerical experiments with dierent weights for
employed one of these three architectures. the ve models. The best combination of weights
Back et al. [8] developed a hybrid architecture was found to be (2, 3, 5, 1 and 4) with prediction
where LDA, logit and GA take care of feature selec- accuracy of 90.78% for data set-I and (1, 5, 4, 2
tion and the selected features were used as predic- and 3) with prediction accuracy of 89.72% for data
tors for BPNN in bankruptcy prediction. The data set-II. Hit ratio of integrated model using equal
consisted of the annual nancial statements of Finn- weights for the output of the ve models was
ish companies. They employed ve hybrid models 89.36%.
viz., (i) DA used for feature selection + DA used Jeng et al. [46] presented a fuzzy inductive learn-
for prediction, (ii) logit used for feature selec- ing method (FILM) that integrated fuzzy set theory
tion + logit used for prediction, (iii) DA used for with regular inductive learning process for predic-
feature selection + BPNN used for prediction, (iv) tion. They compared the FILM with DA and ID3.
logit used for feature selection + BPNN used for Bankruptcy dataset was taken from [68]. The pre-
prediction and (v) GA used for feature selec- diction accuracy of FILM, DA and ID3 was
tion + BPNN used for prediction. They concluded 83.3%, 76.7% and 70%, respectively. Thus they con-
that BPNN and model (v) got improved results than cluded that FILM outperformed both DA and ID3
models (iii) and (iv). Ignizio and Soltys [45] pre- in the case of bankruptcy data. Olmeda and Fernan-
sented a GA-based approach for the simultaneous dez [85] compared the accuracy of classiers in
design and training of neural networks for rm fail- stand-alone mode, on bankruptcy prediction prob-
ure prediction. They called hybrid neural network lem. They also developed a hybrid system integrat-
as ontogenic NN. The data taken from [92] con- ing them. They proposed a framework to
sisted of ve variables. They inferred that the hybrid formulate the choice of the optimal mixture of the
obtained less number of misclassied cases com- technologies as an optimization problem and solved
pared to other methods. Wallrafen et al. [118] stud- it using a genetic algorithm. For the optimal struc-
ied the Genetic algorithm-neural network hybrid ture, they used v-fold cross-validation technique.
22 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

Data for this study was obtained from Spanish set of 71-cases on the CBR component of the system
banking system. For the models combined by a after being pre-classied using SS. Classication
compensating aggregation method they employed values ranged from 16% to 97% for exemption,
the dataset used in Odom and Sharda [82]. BPNN 40% to 93% for preference and 15 to 93% for stay.
performed the best followed by logit, MARS, C4.5 They showed that BanXupport matched cases based
and DA in that order. They combined models in on the selected terms and provided a numerical
two ways: (i) by simple voting scheme and (ii) measure to test the level of similarity to each sec-
by a compensation aggregation method. Overall tion. They concluded that the BanXupport sug-
prediction rates yielded by combined models gested relevant cases with a similarity measure.
viz., NN + logit + C4.5 + DA, NN + logit + C4.5, Ahn et al. [1] proposed hybrid models combining
NN + logit + MARS + DA, NN + logit, NN + lo- rough sets and BPNN for Korean rm failure pre-
git + C4.5 + MARS, NN + logit + DA, NN, all diction and compared it with stand-alone BPNN
methods and NN + logit + MARS by simple voting and DA. They used (i) BPNN (ii) BPNN trained
scheme were 96.21%, 94.94%, 94.69%, 93.93%, with horizontally reduced information system
93.18%, 92.42%, 92.42%, 91.66% and 90.91%, (RNN1) meaning that the feature selection was per-
respectively. They suggested that BPNN model formed before training the NN and (iii) BPNN
was superior to both classical and new statistical trained with horizontally and vertically reduced
and machine learning classiers. They found that information system (RNN2) meaning that feature
hybrid methods by simple voting gave more accu- selection and sample size reduction were carried
rate predictions than the stand-alone methods. out before training NN. The architecture chosen
Gorzalczany and Piasta [36] presented two dier- for BPNN was eight input nodes, ve hidden nodes
ent hybrid intelligent decision support systems viz., and one output node (reported as 8-5-1), while for
(i) neuro-fuzzy classier (N-FC) and (ii) rough clas- RNN1 and RNN2 4-3-1 architecture was used.
sier (RC) for rm bankruptcy prediction. RC was They employed 12-fold cross-validation technique
a combination of rough set based rule induction sys- in testing phase. They constructed two hybrid mod-
tem and statistical techniques. They used Altmans els (i) hybrid model I combining rough sets with
[3] data in this study. Leave-one-out method was RNN1, (ii) hybrid model II combining rough sets
employed to test the models. Classication accura- with RNN2. The average prediction accuracy of
cies of N-FC (with 10 hidden nodes), RC (generated DA, BPNN, RNN1, RNN2, Hybrid I and Hybrid
by ProbRough), C4.5 and CN2 on corporate bank- II were 78.75%, 84.62%, 89.5%, 89.79%, 94.3%
ruptcy data were 97%, 90.4%, 84.8% and 86.4%, and 94.34%, respectively. They inferred that hybrid
respectively. They concluded that N-FC outper- models I and II outperformed both BPNN and
formed other techniques. Elhadi [27] presented a DA.
hybrid system integrating information retrieval Lin and McClean [69] developed hybrid models
(IR) and CBR for legal domain of bankruptcy integrating BPNN, a decision tree (C5.0), DA and
law. He designed an IR-CBR bankruptcy support logistic regression in dierent combinations for pre-
system (BanXupport). Each case was related to a dicting corporate failures in UK. They selected the
bankruptcy law section. BanXupport currently variables using two feature selection methods viz.,
had three clusters corresponding to the three most (i) nancial theory and human judgment (ii)
important sub-sections viz., Stay, exemption and ANOVA. They compared the performance of DA,
preferences. Knowledge in the system was repre- logistic regression, C5.0 and BPNN. Prediction
sented in two ways: IR based and CBR based. Val- accuracy of DA using rst feature selection method
idation of the system was done in two ways: rst was 78.6%, whereas DA with second feature selec-
check the ability of the IR components in selecting tion method produced 77.4%. Prediction accuracy
categories and then check the ability of the CBR of logistic regression using rst feature selection
system in selecting the most similar instances. The method was 84%, whereas logistic regression using
overall accuracy by TDV (term discrimination val- second feature selection method yielded 84.6%. Pre-
ues) an SS (statue-seeded automatic indexing and diction accuracy of BPNN using rst feature selec-
retrieval approach) based on 71-case set was 85% tion method was 87.5%, whereas using BPNN
and 95%, respectively. When using the textbook with second feature selection method yielded
set of 31 cases on the SS component, the system 88.1%. They also proposed some other hybrid mod-
could correctly classify 87% of cases. He ran rst els viz., (i) Hybrid 1 DA + logistic regres-
P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128 23

sion + NN + C5.0, (ii) Hybrid 2 DA + NN + ANN training showed resistance towards overtting
C5.0, (iii) Hybrid 3 logistic regression + C5.0. by keeping the ANN architecture constant. Further,
Using the rst feature selection method the predic- they concluded that GA-based ANN with 1-point
tion accuracy of Hybrids 1, 2 and 3 was 88.4%, crossover, arithmetic crossover and uniform cross-
88.1% and 88.1%, respectively. However, using sec- over performed comparably with BPNN on holdout
ond feature selection method the prediction accu- sample.
racy of Hybrids 1, 2 and 3 was 89.6%, 89.6% and Tung et al. [114] proposed a new neural fuzzy
89.3%. They concluded that ANOVA performed system, viz., the generic self-organizing fuzzy neural
better than human judgment except for DA and network based on the compositional rule of infer-
that the BPNN and C5.0 outperformed statistical ence, GenSoFNN-CRI(S), to predict banking fail-
approaches. They also concluded that hybrid model ure. The interaction between the features was
produced better results. captured in the form of IF_THEN fuzzy rules.
McKee and Lensberg [74] presented a hybrid The GenSoFNN-CRI(S) was compared with the
approach to bankruptcy prediction by integrating Coxs proportional hazards model, BPNN and the
a rough set model with genetic programming modied cerebellar model articulation controller
(GP). They suggested a two-stage hybrid model: (MCMAC). They performed v-fold cross-valida-
stage 1 used a rough set model in Ref. [74] to iden- tion. They concluded that the GenSoFNN-CRI(S)
tify subsets of important explanatory variables and outperformed Coxs model in minimizing type I
stage 2 comprised a GP algorithm in Ref. [59] to error. However, the GenSoFNN-CRI(S) yielded
develop a structural model of bankruptcy based higher type II error rate. They observed that the
on those variables. Rough set model yielded 100% MLP outperformed both the MCMAC and the
classication accuracy on training set and 67% on GenSoFNN-CRI(S). Tseng and Lin [112] proposed
validation set. The GP model obtained 82.6% accu- a hybrid quadratic interval logistic regression anal-
racy on the training set and 80.3% on the validation ysis based on quadratic programming. Their model
set. This was signicantly higher than the accuracy combined the advantages of logistic regression and
achieved by the rough set model. The type I and Tanakas [109] quadratic interval regression model,
type II errors were analyzed using second genetic a variant of fuzzy regression analysis. They obtained
model. New model was trained using the same GP a classication rate of 78%.
algorithm, except that the validation sample of rst
GP model was used as the training data for second 12. Discussion
GP model. These two models yielded 81% and 83%
accuracies respectively over the entire sample. They The following observations can be made from the
concluded that the GP approach coupled with current review paper. Table 2, where the nancial
rough sets could be ecient approach for bank- ratios used in each paper were presented, indicates
ruptcy prediction problem. that a majority of papers considered many nancial
Bian et al. [13] proposed a new fuzzy-rough-near- ratios whereas a few of them considered the Alt-
est neighbour hybrid for bankruptcy prediction. mans variables, which were published in 1968. This
They compared it with crisp K-NN and fuzzy shows that from the domain point of view, the
K-NN. Fuzzy-rough K-NN provided better results bankruptcy research has matured considerably as
in identifying the non-protable companies. When researchers started attaching more importance to
using decision tree feature selection method, fuzzy- other nancial ratios as well. Table 3 presents other
rough K-NN had the lower type I error of 25%. dimensions of the review, wherever available. This
Overall, they concluded that fuzzy-rough K-NN table indicates that a vast majority of the studies
performed better in minimizing type I error, while reviewed here pertain to rms and not banks. Fur-
having satisfactory type II error. Pendharkar and ther, majority of the studies are conducted by taking
Rodger [88] studied GA-based ANN on bankruptcy the data sets relevant to the time period 19802003.
prediction problem. They chose arithmetic cross- Moreover, for a vast majority of the works, the
over operator for the GA-based ANN. They used country of origin of the data sets is USA followed
real-valued GA to learn the connection weights in by European countries. Further, it is observed that
an ANN. They used two datasets of which the rst a variety of statistical and intelligent techniques
one is taken from [49] and second data set was a have already been applied to the bankruptcy predic-
real-life data set. They showed that GA-based tion problem.
24 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

The general observation is that the statistical bankruptcy prediction research. The interest and
techniques such as logistic regression, LDA, QDA, condence in techniques such as fuzzy rule based
FA were all outperformed by the most popular classiers, SVM, rough sets, isotonic separation
NN architecture viz., BPNN, wherever comparisons etc has grown enormously during the past 5
were made between these two families of techniques. 10 years. Statistical techniques are no longer pre-
This is not surprising because the BPNN with logis- ferred in view of their low accuracy. Further,
tic activation function can be thought of as a conu- the BPNN is also very much exploited.
ence of several logistic regressions tted together in Other NN architectures deserve much wider appli-
parallel. Thus, the non-linearity in the data can be cation in this eld. However, decision trees such
modeled better by the BPNN. Comparing decision as CART are not employed as much as they
trees with NN architectures, even in cases where deserve.
they yield identical performance, we recommend
the use of decision trees as they produce an impor- 13. Conclusions and future directions
tant by-product viz., ifthen rules. These rules can
be used as an early warning expert system later on. A comprehensive review is conducted on the
Depending on the data set, both techniques are research work done during 19682005 in the appli-
capable of outperforming each other in terms of cation of statistical and intelligent techniques to
accuracy. However, in the case of bankruptcy pre- solve the bankruptcy prediction problem in banks
diction problems, BPNN outperformed the decision and rms. The review is organized by taking the
trees such as C4.5 and ID3 in terms of accuracy. type of technique as an important dimension
Further, BPNN, DA and logistic regression outper- besides others such as the source of data sets and
formed the CBR implemented without weights. variables used and the comparative performance
Thus, CBR with the simple k-nearest neighbor of techniques in terms of prediction accuracy, wher-
method at its heart cannot solve classication prob- ever available. An important conclusion that can be
lems with nonlinear boundaries. Also, CBR cannot drawn from the review is that researchers employed
generalize well. almost all intelligent techniques to solve this impor-
Then, rough set based approach outperformed tant problem. The review indicates that statistical
DA, logistic regression and a decision tree. How- techniques in stand-alone mode are no longer
ever, rough set based approaches can be, in gen- employed and among the stand-alone intelligent
eral, inaccurate and sensitive to changes in data. techniques, neural networks were the most often
As regards DEA, it yields us only the relative used family followed by rough sets, CBR, OR tech-
scores of the eciency levels of the banks. On niques, evolutionary approaches and other tech-
the other hand, SVM can be very accurate and niques subsuming fuzzy logic, SVM, etc. An
does not have local minima problems unlike important trend has been to build soft computing
BPNN and it can also be trained with a small architectures (hybrid intelligent systems) for the
training set. It is observed that SVM outperformed problem. This review supports the popular opinion
DA, logistic regression and BPNN. Fuzzy logic that ensemble classiers, by and large, outperform
based techniques are least exploited in bankruptcy the individual techniques, that constitute the ensem-
prediction research. In the few studies where they ble classiers. As regards future directions, a lot of
were applied, fuzzy rule based classier outper- scope still exists in developing new hybrid systems
formed LDA and logistic regression, but BPNN in various forms for the problem. The full potential
outscored fuzzy rule based classier. More impor- of intelligent techniques such as SVM, fuzzy logic,
tantly, fuzzy ifthen rules can be used as early collaborative ltering and hitherto unused neural
warning fuzzy expert systems. In general, in the network architectures such as radial basis function
soft computing architectures, ensemble classiers network, reduced coulomb energy network, gener-
outperformed the individual models, which is on alized regression neural network, counter propaga-
the expected lines. Further, it can be observed that tion neural network, adaptive resonance theory
the interest to design and employ a variety of soft (ART) network and decision trees such as CART,
computing architectures has increased signicantly TreeNet, random forests can be exploited. Given
over the past decade. the limitations of all intelligent techniques in
In view of the foregoing and from Tables 2 and stand-alone mode, the authors feel that it is worth-
3, the authors noticed a signicant trend in the while to investigate newer soft computing architec-
P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128 25

tures (hybrid intelligent systems) in order to amplify International Journal of Intelligent Systems in Accounting,
the advantages of the individual models and mini- Finance and Management 6 (1997) 249264.
[13] H. Bian, L. Mazlack, Fuzzy-rough nearest neighbour
mize their limitations. One such hybrid system classication approach, in: 22nd International Conference
could focus on exploiting the enormous domain of the North American Fuzzy Information Processing
knowledge that is extant for this problem besides Society (NAFIPS 2003) Proceedings Chicago, 2003, pp.
using data-driven modeling. The rst module of 500505.
such a hybrid could use the nancial domain [14] J.C. Bioch, V. Popova, Bankruptcy prediction with rough
sets, ERIM Report Series Research in Management (ERS-
experts experiential knowledge to build a fuzzy 2001-11-LIS), 2001.
inference system while the second module could [15] E. Brewer, W. Jackson, J. Moser, Alligators in the Swamp:
be any of the intelligent techniques such as neural The impact of derivatives on the nancial performance of
networks, decision trees, etc. depository institutions, Journal of Money, Credit and
Banking 28 (1996) 478496.
[16] S.M. Bryant, A case-based reasoning approach to bank-
References ruptcy prediction modeling, Intelligent Systems in Account-
ing, Finance and Management 6 (1997) 195214.
[1] B.A. Ahn, S.S. Cho, C.Y. Kim, The integrated methodol- [17] T.G. Calderon, J.J. Cheh, A Roadmap for future neural
ogy of rough set theory and articial neural network for networks research in auditing and risk assessment, Inter-
business failure prediction, Expert Systems with Applica- national Journal of Accounting Information Systems 3
tions 18 (2000) 6574. (2002) 203236.
[2] P. Alam, D. Booth, K. Lee, T. Thordarson, The use of [18] S. Canbas, A. Cabuk, S.B. Kilic, Prediction of commercial
fuzzy clustering algorithm and self-organizing neural net- bank failure via multivariate statistical analysis of nancial
work for identifying potentially failing banks: An experi- structure: The Turkish case, European Journal of Opera-
ment study, Expert Systems with Applications 18 (2000) tional Research 166 (2005) 528546.
185199. [19] A. Cielen, L. Peeters, K. Vanhoof, Bankruptcy prediction
[3] E.I. Altman, Financial ratios, discriminant analysis and the using a data envelopment analysis, European Journal of
prediction of corporate bankruptcy, Journal of Finance 23 Operational Research 154 (2004) 526532.
(1968) 589609. [20] R. Cole, J. Gunther, A CAMEL ratings shelf life, Federal
[4] E.I. Altman, G. Marco, F. Varetto, corporate distress Reserve Bank of Dallas Review (1995) 1320.
diagnosis: Comparisons using linear discriminant analysis [21] C.A. Cooper, E.W.W. Rhodes, Measuring the eciency of
and neural networks (the Italian experience), Journal of decision making units, European Journal of Operational
Banking and Finance 18 (1994) 505529. Research 2 (1978) 429444.
[5] E.I. Altman, R.G. Haldeman, P. Narayanan, ZETA [22] N. Cristiamini, J. Shawe-Taylor, An Introduction to
ANALYSIS, a new model to identify bankruptcy risk of Support Vector Machines, Cambridge university press,
corporations, Journal of Banking and Finance 1 (1977) 29 Cambridge, England, 2000.
54. [23] M. Daubie, N. Meskens, Business failure prediction: A
[6] J.D. Andres, M. Landajo, P. Lorca, Forecasting business review and analysis of the literature. Working Paper,
protability by using classication techniques: A compar- Department of Productions and Operations Management,
ative analysis based on a Spanish case, European Journal of Catholic University of Mons, Belgium, 2002, pp. 115.
Operational Research 167 (2005) 518542. [24] J.R. Dietrich, R.S. Kaplan, Empirical analysis of the loan
[7] A.F. Atiya, Bankruptcy prediction for credit risk using classication decision, The Accounting Review 57 (1982)
neural networks: A survey and new results, IEEE Trans- 1838.
actions on Neural Networks 12 (4) (2001) 929935. [25] A.I. Dimitras, S.H. Zanakis, C. Zopounidis, A survey of
[8] B. Back, T. Laitinen, K. Sere, Neural network and genetic business failure with an emphasis on prediction method and
algorithm for bankruptcy prediction, Expert Systems with industrial applications, European Journal of Operational
Applications 11 (4) (1996) 407413. Research 90 (1996) 487513.
[9] J. Baek and S. Cho, Bankruptcy Prediction for credit risk [26] A.I. Dimitras, R. Slowinski, R. Susmaga, C. Zopounidis,
using an auto associative neural networks in Korean rms, Business failure prediction using rough sets, European
in: IEEE International Conference on Computational Journal of Operational Research 114 (1999) 263280.
Intelligence for Financial Engineering, Hong-Kong, 2003. [27] M.T. Elhadi, Bankruptcy support system: Taking advan-
[10] W.J. Banks, L.A. Prakash, On the performance of linear tage of information retrieval and case-based reasoning,
programming heuristics applied on a quadratic transfor- Export Systems with Applications 18 (2000) 215219.
mation in the classication problem, European Journal of [28] S.E. Fahlman, C. Lebiere, The cascade-correlation learning
Operational Research 74 (23) (1994) 2328. architecture, in: Advances in Neural Information Process-
[11] R. Barniv, A. Anurag, R. Leach, Predicting the out come ing Systems, Morgan Kaufmann, San Mated, CA, 1990, pp.
following bankruptcy ling: A three state classication 524532.
using NN, International Journal of Intelligent Systems in [29] D. Fletcher, E. Goss, Application forecasting with neural
Accounting, Finance and Management 6 (1997) 177 networks an application using bankruptcy data, Informa-
194. tion and Management 24 (1993) 159167.
[12] T.B. Bell, Neural nets or the logit model? A comparison of [30] D. Fraser, The determinants of bank prots: An analysis of
each models ability to predict commercial bank failures, extremes, Financial Review 11 (1976) 6987.
26 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

[31] H. Frydman, E.I. Altman, D. Kao, Introducing recursive tion, Expert Systems with Applications 11 (4) (1996) 415
partitioning for nancial classication: The case of nancial 422.
distress, Journal of Finance 40 (1) (1985) 269291. [49] E.A. Joachimsthaler, A. Stam, Four approaches to the
[32] R. Gady, Anatomy of protable medium-size banks in the classication problem in discriminant analysis: An experi-
fourth district, 19661970, Economic Review, Federal mental study, Decision Sciences 19 (1988) 322333.
Reserve Bank of Cleveland, 1972, pp. 2032. [50] R.A. Johnson, D.W. Wichern, Applied Multivariate Sta-
[33] J. Gehrke, R. Ramakrishnan, W.Y. Loh, BOAT-optimistic tistical Analysis, second ed., Prentice-Hall, Englewood
decision tree construction, in: Proceedings ACM SIGMOD Clis, NJ, 1988.
International Conference Management of Data, Philadel- [51] S. Jones, D.A. Hensher, Predicting rm nancial distress: A
phia, PA, 1999, pp. 169180. mixed logit model, Accounting Review 79 (4) (2004) 1011
[34] J. Gentry, A. Newbold, D. Whitford, Classifying bankrupt 1038.
rms with funds ow components, Journal of Accounting [52] C. Kao, S.-T. Liu, Prediction bank performance with
Research 23 (11(1)) (1985) 146160. nancial forecasts: A case of Taiwan commercial banks,
[35] A. Gersho, R.M. Gray, Vector Quantization and Signal Journal of Banking & Finance 28 (2004) 23532368.
Compression, Kluwer, Norwell, MA, 1992. [53] G.V. Karels, A.J. Prakash, Multivariate normality and
[36] M.B. Gorzalczany, Z. Pista, Neuro-fuzzy approach versus forecasting of business bankruptcy, Journal of Business
rough-set inspired methodology for intelligent decision Finance and Accounting 14 (4) (1987).
support, Information Science 120 (1999) 4568. [54] S. Kaski, J. Sinkkonen, J. Peltonen, Bankruptcy analysis
[37] S. Greco, B. Matarazzo, R. Slowinski, A new rough set with self-organizing maps in learning metrics, IEEE Trans-
approach to evaluation of bankruptcy risk, in: C. Zopo- actions on Neural Networks 12 (4) (2001).
unidis (Ed.), Operational Tools in the Management of [55] K. Kiviluoto, Predicting bankruptcies with self organizing
Financial Risks, Kluwer Academic Publishers, Dordrecht, map, Neurocomputing 21 (1998) 191201.
1998, pp. 121136. [56] T. Kohonen, Self-Organizing Maps, second ed., Springer-
[38] S. Greco, B. Matarazzo, R. Slowinski, A new rough set Verlag, Berlin, 1997.
approach to multicriteria and multiattribute classication, [57] J. Kolari, D. Glennon, H. Shin, M. Caputo, Predicting
Rough Sets and Current Trends in Computing, Berlin large US commercial bank failures, Journal of Economics
(1998) 6067. and Business 54 (32 1) (2002) 361387.
[39] S. Greco, B. Matarazzo, R. Slowinski, Decision rule [58] J.L. Kolodner, Case Based Reasoning, Morgan Kaufman,
approach, in: J. Figueira, S. Greco, M. Ehrgott (Eds.), CA, 1993.
Multiple Criteria Decision Analysis: State of the Art Surveys, [59] J.R. Koza, Genetic Programming: On the Programming of
Springer-Verlag, New York, 2005, pp. 507562 (Chapter 13). Computers by Means of Natural Selection, Massachusetts
[40] S. Greco, B. Matarazzo, R. Slowinski, Rough sets meth- Institute of Technology, Cambridge, MA, 1992.
odology for sorting problems in presence of multiple [60] M. Kwast, J. Rose, Pricing, Operating Eciency and
attributes and criteria, European Journal of Operational Protability Among Large Commercial Banks, Journal of
Research 138 (2) (2002) 247259. Banking and Finance 6 (1982) 233254.
[41] S. Greco, B. Matarazzo, R. Slowinski, Rough approxima- [61] P.A. Lachenbruch, An almost unbiased method of obtain-
tion by dominance relations, International Journal of ing condence intervals for the probability of misclassi-
Intelligent Systems 17 (2) (2002) 153171. cation in discriminant analysis, Biometrics (December)
[42] S. Greco, B. Matarazzo, R. Slowinski, Multicriteria clas- (1967) 639645.
sication, in: W. Kloesgen, J. Zytkow (Eds.), Handbook of [62] R.C. Lacher, P.K. Coats, S.C. Sharma, L.F. Fantc, A
Data Mining and Knowledge Discovery, Oxford University Neural network for classifying the nancial health of a rm,
Press, New York, 2002, pp. 318328 (Chapter 16.1.9). European Journal of Operational Research 85 (1995) 5365.
[43] S. Greco, B. Matarazzo, R. Slowinski, Rough sets theory [63] K.F. Lam, J.W. Moy, Combining discriminant methods in
for multicriteria decision analysis, European Journal of solving classication problems in two-group discriminant
Operational Research 129 (1) (2001) 147. analysis, European Journal of Operational Research 138
[44] J. Haslem, A. Scheraga, C.A. Bedingeld, P. James, An (2002) 294301.
Analysis of the Foreign and Domestic Balance Sheet [64] M. Lam, Neural networks techniques for nancial perfor-
Strategies of the U.S. Banks and Their Association to mance prediction: integrating fundamental and technical
Protability Performance, Management International analysis, Decision Support Systems 37 (2004) 567581.
Review. First Quarter, Wiesbaden, 1992. [65] K.C. Lee, I. Han, Y. Kwon, Hybrid neural network models
[45] J.P. Ignizio, J.R. Soltyas, Simultaneous design and training for bankruptcy predictions, Decision Support Systems 18
of ontogenic neural network classier, Computers Opera- (1996) 6372.
tions Research 23 (6) (1996) 535546. [66] K. Lee, D. Booth, P. Alam, A comparison of supervised
[46] B. Jeng, Y.-M. Jeng, T.-P. Liang, FILM: A fuzzy learning and unsupervised neural networks in predicting bankruptcy
method for automated knowledge acquisition, Decision of Korean rms, Expert Systems with Applications 29
Support Systems 21 (1997) 6173. (2005) 116.
[47] H. Jo, I. Han, H. lee, Bankruptcy prediction using case- [67] M. Leshno, Y. Spector, Neural network prediction analysis:
based reasoning, neural network and discriminant analysis The bankruptcy case, Neurocomputing 10 (1996) 125
for bankruptcy prediction, Expert Systems with Applica- 147.
tions 13 (2) (1997) 97108. [68] T.P. Liang, A composite approach to inducing knowledge
[48] H. Jo, I. Han, Integration of case-based forecasting, neural for expert systems design, Management Science 38 (1)
network and discriminant analysis for bankruptcy predic- (1992) 145154.
P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128 27

[69] F.Y. Lin, S. McClean, A datamining approach to the ruptcy prediction, Expert Systems with Applications 23 (3)
prediction of corporate failure, Knowledge Based Systems (2002) 255264.
14 (2001) 189195. [87] Z. Pawlak, Rough sets, International Journal of Computer
[70] M.L. Marais, J. Patel, M. Wolfson, The experimental and Information Science 11 (1982) 341356.
design of classication models: An application of recursive [88] P.C. Pendharkar, J.A. Rodger, An empirical study of
partitioning and bootstrapping to commercial bank loan impact of crossover operators on the performance of non-
classications, Journal of Accounting Research 22 (1984) binary genetic algorithm based neural approaches for
87113. classication, Computers and Operations Research 31
[71] D. Martin, Early warning of bank failure: A logit regression (2004) 481498.
approach, Journal of Banking and Finance 1 (1977) 249 [89] S. Piramuthu, H. Ragavan, M.J. Shaw, Using feature
276. construction to improve the performance of the neural
[72] W.S. McCulloch, W.A. Pitts, Logical calculus of ideas networks, Management Science 44 (3) (1998).
immanent in nervous activity, Bulletin of Mathematical [90] H.D. Platt, M.B. Platt, J. Pedersen, Bankruptcy discrimi-
Biophysics 5 (1943) 115133. nation with real variables, Journal of Business Finance and
[73] T.E. McKee, Developing a bankruptcy prediction model Accounting 21 (1994) 491509.
via rough sets theory, International Journal of Intelligent [91] J.R. Quinlan, Decision trees as probabilistic classiers, in:
Systems in Accounting, Finance and Management 9 (2000) Proceedings of 4th International Workshop Machine
59173. Learning, Irvine, CA, 1987, pp. 3137.
[74] T.E. McKee, T. Lensberg, Genetic programming and rough [92] E. Rahimian, S. Singh, T. Thammachote, R. Virmani,
sets: A hybrid approach to bankruptcy classication, Bankruptcy prediction by neural network, in: R.R. Trippi,
European Journal of Operational Research 138 (2002) E. Turban (Eds.), Neural Networks in Finance and
436451. Investing, Irwin Professional Publishing, Burr Ridge,
[75] T.E. McKee, Rough sets bankruptcy prediction models USA, 1996.
versus auditor signaling rates, Journal of Forecasting 22 [93] A.-K. Rashad, K.M. El-Sheshai, Information choice and
(2003) 569589. utilization in an experiment on default prediction, Journal
[76] R. Merton, On the pricing of corporate debt: The risk of Accounting Research 18 (2) (1980).
structure of interest rates, Journal of Finance 29 (1974) [94] O.P. Rud, Data Mining Cook Book, John Wiely and Sons,
449470. New York, 2000.
[77] W.F. Messier, J. Hansen, Inducing rules for expert system [95] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning
development: An example using default and bankruptcy internal representations by error propagation, in: D.E.
data, Management Science 12 (1988) 14031415. Rumelhart, J.L. McClelland (Eds.), Parallel Distributed
[78] S. Michael, D. Georgios, M. Nikolaos, Z. Constantin, A Processing: Explorations in the Microstructure of Cogni-
fuzzy knowledge-based decision aiding method for the tion, vol. 1, The MIT Press, Cambridge, MA, 1986, pp.
assessment of nancial risk: The case of corporate bank- 318362.
ruptcy prediction, in: European Symposium on Intelligent [96] S. Russell, P. Norvig, Articial Intelligence. A Modern
Techniques, Crete, Greece, 1999. Approach, second ed., Prentice-Hall, New Jersey, USA,
[79] J.H. Min, Y.-C. Lee, Bankruptcy prediction using support 2003.
vector machine (SVM) with optimal choice of kernel [97] Y.U. Ryu, W.T. Yue, Firm bankruptcy prediction: Exper-
function parameters, Expert Systems with Applications 28 imental comparison of isotonic separation and other
(2005) 603614. classication approaches, IEEE Transactions On Systems,
[80] S. Nanda, P. Pendharkar, Linear models for minimizing Management and Cybernetics-Part A: Systems and
misclassication costs in bankruptcy prediction, Interna- Humans 35 (5) (2005) 727737.
tional Journal of Intelligent Systems in Accounting, [98] L. Salchenberger, C. Mine, N. Lash, Neural networks: A
Finance and Management 10 (2001) 155168. tool for predicting thrift failures, Decision Sciences 23
[81] J. Moody, C.J. Darken, Fast learning in networks of (1992) 899916.
locally-tuned processing units, Neural Computing 1 (1989) [99] J. Scott, The probability of bankruptcy, Journal of Banking
281294. and Finance 5 (1981) 317344.
[82] M. Odom, R. Sharda, A neural network for bankruptcy [100] C. Serrano-Cinca, Self organizing neural networks for
prediction, in: IJCNN International Conference on Neural nancial diagnosis, Decision Support Systems 17 (1996)
Networks, San Diego, CA, 1990. 227238.
[83] J.A. Ohlson, Financial rations and the probabilistic predic- [101] R. Sharda, R.L. Wilson, Performance comparison issues in
tion of bankruptcy, Journal of Accounting Research 18 neural network experiments for classication problems, in:
(1980) 109131. Proceedings of the 26th Hawai International Conference on
[84] D.E. OLeary, Using neural network to predict corporate System Scientists, 1993.
failure, International Journal of Intelligent Systems in [102] K.-S. Shin, Y.-J. Lee, A genetic algorithm application in
Accounting Finance and Management 7 (1998) 187197. bankruptcy prediction modeling, Expert Systems with
[85] I. Olmeda, E. Fernandez, Hybrid classiers for nancial Applications 23 (3) (2002) 321328.
multicriteria decision making: The case of bankruptcy [103] R. Slowinski, S. Greco, B. Matarazzo, Rough set based
prediction, Computational Economics 10 (1997) 317 decision support, in: E.K. Burke, G. Kendall (Eds.), Search
335. Methodologies: Introductory Tutorials in Optimization and
[86] C.-S. Park, I. Han, A case-based reasoning with the feature Decision Support Techniques, Springer-Verlag, New York,
weights derived by analytic hierarchy process for bank- 2005, pp. 475527 (Chapter 16).
28 P. Ravi Kumar, V. Ravi / European Journal of Operational Research 180 (2007) 128

[104] D.F. Specht, Probabilistic neural networks, Neural Net- [117] L. Wall, Why are some banks more protable than others?
works 3 (1990) 110118. Journal of Bank Research 15 (1985) 240256.
[105] T.K. Sung, N. Chank, G. Lee, Dynamics of modelling in [118] J. Wallrafen, P. Protzel, H. Popp, Genetically optimized
data mining: Interpretive approach to bankruptcy predic- neural network classiers for bankruptcy prediction, in:
tion, Journal of Management Information Systems 16 Proceedings of the 29 the Annual Hawaii International
(1999) 6385. Conference on System Sciences, 1996.
[106] P. Swicegood, J.A. Clark, O-site monitoring for predicting [119] R.C. West, A factor analytic approach to bank condition,
bank under performance: A comparison of neural net- Journal of Banking and Finance 9 (1985) 253266.
works, discriminant analysis and professional human judg- [120] J.W. Wilcox, A prediction of business failure using account-
ment, International Journal of Intelligent Systems in ing data, empirical research in accounting: Selected studies,
Accounting, Finance and Management 10 (2001) 169186. Journal of Accounting Research (Suppl.) (1973) 163
[107] K.Y. Tam, Neural network models and the prediction of 179.
bank bankruptcy, Omega 19 (5) (1991) 429445. [121] R.L. Wilson, R. Sharda, Bankruptcy prediction using
[108] K.Y. Tam, M. Kiang, Predicting bank failures: A neural neural networks, Decision Support Systems 11 (1994)
network approach, Decision Sciences 23 (1992) 926947. 545557.
[109] H. Tanaka, H. Lee, Interval regression analysis by qua- [122] Z.R. Yang, M.B. Platt, H.D. Platt, Probability neural
dratic programming approach, IEEE Transactions on network in bankruptcy prediction, Journal of Business
Fuzzy Systems 6 (4) (1998) 473481. Research 44 (1999) 6774.
[110] F.E.H. Tay, L. Shen, Economic and nancial prediction [123] A.Y.N. Yip, Predicting business failure with a case-based
using rough set model, European Journal of Operational reasoning approach, lecture notes in computer science, in:
Research 141 (2002) 641659. M.G. Negoita, R.J. Howlett, L.C. Jain (Eds.), Knowledge-
[111] R.R. Trippi, E. Turban (Eds.), Neural Networks in Finance Based Intelligent Information and Engineering Systems: 8th
and Investing, Probus Publishing, Chicago, 1993 (Chapters International Conference, KES 2004, Wellington, New
813). Zealand, September 3215/2004, Proceedings, Part III,
[112] F.-M. Tseng, L. Lin, A quadratic interval logit model for 2004, pp. 2025.
forecasting bankruptcy, Omega 33 (2005) 8591. [124] L.A. Zadeh, Fuzzy sets, Information and Control 8 (1965)
[113] J. Tsukuda, S.-I. Baba, Predicting Japanese corporate 338353.
bankruptcy in terms of nance data using neural network, [125] L.A. Zadeh, Soft computing and fuzzy logic, IEEE Soft-
Computers and industrial Engineering 27 (14) (1994) 445 ware 11 (6) (1994) 4856.
448. [126] G. Zhang, M.Y. Hu, B.E. Patuwo, D.C. Indro, Articial
[114] W.L. Tung, C. Quek, P. Cheng, GenSo-EWS: A novel neural networks and bankruptcy prediction general frame-
neural-fuzzy based early warning system for predicting work and cross-validated analysis, European Journal of
bank failures, Neural Networks 17 (2004) 567587. Operational Research 116 (1999) 1632.
[115] V. Vapnik, in: S. Haykin (Ed.), Statistical Learning Theory. [127] H.J. Zimmermann, Fuzzy set theory and its applications,
Adaptive and Learning Systems, vol. 736, John Wiley and Kluwer Academic Publishers, London, 1996.
Sons, 1998. [128] M. Zmijewski, Methodological issues related to the estima-
[116] F. Varetto, Genetic algorithm applications in the analysis of tion of nancial distress prediction models, Journal of
insolvency risk, Journal of Banking and Finance 22 (1998) Accounting Research 22 (1984) 5982.
14211439.

Você também pode gostar