Abstract

This study aimed to develop and compare statistical models using the techniques of artificial neural networks (ANNs) and support vector machines (SVMs) to investigate which one offers the best results in evaluating insolvency of mutual credit unions. The information required to build the models were obtained with a sample of 62 credit unions (31 solvent and 31 insolvent) to which financial indicators of the PEARLS (Protection, Effective financial structure, Asset quality, Rates of return and cost, Liquidity and Signs of growth) system were calculated. The RBF network, multilayer perceptron, multilayer perceptronCS and LibSVM algorithms were used to obtain the ANNs and SVMs; for each algorithm, the ANNs were built with three groups of indicators (27, 11 and 10 indicators). This is the first study done with ANNs in Brazilian credit unions. When analyzing the results of ANNs and SVMs, the superiority of the SVMs as binary classifier for evaluating insolvency was evidenced, since its LibSVM algorithm showed the best results in all assessments of performance proposed in this study. The only LibSVM indicator with performance inferior to ANNs was the error rate of the negative class which indicates those negative class data that were classified incorrectly.

Key words: Insolvency, credit unions, artificial neural networks, support vector machines.