Improving the performance of statistical learning methods with a combined meta-heuristic for consumer credit risk assessment

In credit risk applications, it is common to use statistical learning methods for classification. It is possible to enhance accuracy of these methods by identifying an optimum feature subset of variables instead of whole set. In this study, a combination of two well-known meta-heuristics was used together with a statistical method on the basis of two stages. In the first stage, prediction accuracy performances of several statistical learning methods on a real world dataset from UCI database were compared with Friedman rank sum test. In the second stage, a promising method was chosen from the previous stage and implemented into the combined meta-heuristic. Subset selection was performed by the heuristic and classifier learned with selected variables on the same dataset from the first stage. Performance improvement was validated with Wilcoxon rank-sum test. Promising results were reached compared to that presented in the literature, and credit assessment and scoring practitioners may highly benefit from the proposed approach.