2. Predicting 5nancial distress: neural network models
American Express is using a neural network-based system to detect incidences of credit card
fraud, and Lloyds Bowmaker Motor Finance has deployed a neural network credit scoring system
for automobile "nancing decisions. Lloyds claims that the neural-based credit scoring system is
10% more accurate than the system it replaced [18]. Security Paci"c Bank (SPB) is also using
a neural network intelligent system for credit scoring of small business loans [18]. The speci"c
neural network credit scoring model developed by SPB is a multi-layer perceptron (MLP) trained
by the back-propagation learning algorithm. SPB believes the advantage of the neural network
scoring system is the improved function-"tting capability due to the intrinsic non-linear pattern
recognition capability of the neural network. They state that even a small improvement in
predictive accuracy of the neural network credit scoring model is critical; a 1% improvement in
accuracy will reduce losses in a large loan portfolio and save millions of dollars.
Due to the proprietary nature of credit scoring, there is a paucity of research reporting the
performance of commercial credit scoring applications. The research that exists today focuses on
two areas, the prediction of "rm insolvency and the prediction of individual credit risk. Research at
the "rm level is reported "rst, followed by research at the level of the individual consumer. Altman
et al. employs both linear discriminant analysis and a multilayer perceptron neural network to
diagnose corporate "nancial distress for 1000 Italian "rms [8]. The authors conclude that neural
networks are not a clearly dominant mathematical technique compared to traditional statistical
techniques such as discriminant analysis, and that linear discriminant analysis compares rather
well to the neural network model in decision accuracy [8]. Coats and Fant report a di!erent
experience, contrasting a multilayer perceptron neural network with linear discriminant analysis
for a set of "rms labeled viable or distressed obtained from COMPUSTAT for the period
1970}1989 [11]. They "nd the neural network to be more accurate than linear discriminant
analysis, particularly for predicting "rms in "nancial distress. Lacher et al. utilizes the Coats and
Fant data to predict "nancial distress with Altman's Z score [15]. They state that the neural
network developed by cascade correlation more accurately predicts the "nancial health of a "rm
[15]. Salchenberger et al. investigate the use of a multilayer perceptron neural network to predict
the "nancial health of savings and loans [17]. The authors compare a multilayer perceptron neural
network with a logistic regression model for a data set of 3429 S&L's from January 1986 to
December 1987. They "nd that the neural network model performs as well as or better than the
logistic regression model for each data set examined. Tam and Kiang studied the application of
a multilayer perceptron neural network model to Texas bank failure prediction for the period
1985}1987 [7]. The neural network prediction accuracy is compared to linear discriminant
analysis, logistic regression, k nearest neighbor, and a decision tree model. Their results suggest
that the multilayer perceptron is most accurate, followed by linear discriminant analysis, logistic
regression, decision trees, and k nearest neighbor.
Desai et al. [13] investigate a multilayer perceptron neural network, a mixture of experts neural
network, linear discriminant analysis, and logistic regression for scoring credit applicants in the
credit union industry. Their methodology consists of two-fold cross validation of "eld data
obtained from three credit unions and the assumption of equal costs for good and bad credit risks.
They conclude that the neural network models outperform linear discriminant analysis, but are
only marginally better than logistic regression models. The authors also report that the neural