Artificial Neural Networks to Evaluate Bank Financial Condition -- a Comparison of Relative Costs of Misclassification

This study compares the financial distress classification accuracy of two traditional analytical techniques for evaluating financial distress models: logit analysis and multiple discriminant analysis (MDA) to the less restrictive and non-parametric computer-aided techniques, Artificial Neural Networks (ANNs). The study evaluates the techniques both for their overall classification accuracy and type I and type II misclassification rates. Five different ANNs were used because of their suitability for classification-type problems. The techniques were tested on select financial ratios on four areas of a bank's financial performance on 994 financially healthy and 145 financially distressed banks. The results indicated that logit and MDA to have lower overall error rates than ANNs. However, when the relative costs of misclassification were compared, ANNs, in general, performed better than logit and MDA