Understanding Bank Failure: A Close Examination of Rules Created by Genetic Programming

This paper presents a novel method to predict bankruptcy, using a Genetic Programming (GP) based approach called Evolving Decision Rules (EDR). In order to obtain the optimum parameters of the classifying mechanism, we use a data set, obtained from the US Federal Deposit Insurance Corporation (FDIC). The set consists of limited financial institutions' data, presented as variables widely used to detect bank failure. The outcome is a set of comprehensible decision rules, which allows to identify cases of bankruptcy. Further, the reliability of those rules is measured in terms of the true and false positive rate, calculated over the whole data set and plot over the Receiving Operating Characteristic (ROC) space. In order to test the accuracy performance of the mechanism, we elaborate two experiments: the first, aimed to test the degree of the variables' usefulness, provides a quantitative and a qualitative analysis. The second experiment completed over 1000 different re-sampled cases is used to measure the performance of the approach. To our knowledge this is the first computational technique in this field able to give useful insights of the method's predictive structure. The main contributions of this work are three: first, we want to bring to the arena of bankruptcy prediction a competitive novel method which in pure performance terms is comparable to state of the art methods recently proposed in similar works, second, this method provides the additional advantage of transparency as the generated rules are fully interpretable in terms of simple financial ratios, third and final, the proposed method includes cutting edge techniques to handle highly unbalanced samples, something that is very common in bankruptcy applications.