Predicting regional credit ratings using ensemble classification with metacost

Abstract

Ensemble classifiers are learning algorithms that combine sets of base classifiers in order to increase their diversity and, thus, decrease variance and achieve better predictive performance compared to single classifiers. Previous research has shown that ensemble classifiers are more accurate than single classifiers in predicting credit ratings. Here we deal with highly imbalanced multi-class data of regional entities. To overcome these problems, we propose a novel hybrid model combining data oversampling and cost-sensitive ensemble classification. This paper demonstrates that the use of the SMOTE technique to balance the multi-class data solves the imbalance problem effectively. Different misclassification cost assigned in cost matrix solves the problem of ordered classes. This approach is combined with ensemble classification within the MetaCost framework. We show that more accurate prediction can be achieved using this approach in terms of average cost and area under ROC. This paper provides empirical evidence on the dataset of 451 regions classified into 8 rating classes, as obtained from the Moody’s rating agency. The results show that Random Forest combined with MetaCost outperforms the rest of the base classifiers, as well as other benchmark methods.