An exploratory analysis of methods for extracting credit risk rules

Abstract

This paper performs a comparative analysis of two kind of methods for extracting credit risk rules. On one hand we have a set of methods based on the combination of an optimization technique initialized with a neural network. On the other hand there are partition algorithms, based on trees. We show results obtain on two real databases. The main findings are that the set of rules obtained by the first set of methods give a set of rules with a reduced cardinality, with an acceptable precision regarding classification. This is a desirable property for financial institutions, who want to decide credit approval face to face with customers. Bank employees who daily deal with retail customers can be easily trained for selecting the best customers, by using this kind of solutions.