Neural network credit scoring models

Abstract

This paper investigates the credit scoring accuracy of five neural network models: multilayer perceptron, mixture-of-experts, radial basis function, learning vector quantization, and fuzzy adaptive resonance. The neural network credit scoring models are tested using 10-fold crossvalidation with two real world data sets. Results are benchmarked against more traditional methods under consideration for commercial applications including linear discriminant analysis, logistic regression, k nearest neighbor, kernel density estimation, and decision trees. Results demonstrate that the multilayer perceptron may not be the most accurate neural network model, and that both the mixture-of-experts and radial basis function neural network models should be considered for credit scoring applications. Logistic regression is found to be the most accurate of the traditional methods. Scope and purpose In the last few decades quantitative methods known as credit scoring models have been developed for the credit granting decision. The objective of quantitative credit scoring models is to assign credit applicants to one of two groups: a “good credit” group that is likely to repay the financial obligation, or a “bad credit” group that should be denied credit because of a high likelihood of defaulting on the financial obligation. The first model employed for credit scoring, and a commonly used method today, is linear discriminant analysis, a simple parametric statistical method. With the growth of the credit industry and the large loan portfolios under management today, the industry is actively developing more accurate credit scoring models. Even a fraction of a percent increase in credit scoring accuracy is a significant accomplishment. This effort is leading to the investigation of nonparametric statistical methods, classification trees, and neural network technology for credit scoring applications. The purpose of this research is to investigate the accuracy of five neural network architectures for the credit scoring applications and to benchmark their performance against the models currently under investigation today.