One-class classification for credit scoring

Abstract

Credit scoring nowadays has become a crucial tool for the bank industry since its purpose is to assess the risk associated with granting clients loans. Accurate prediction of credit risk according to historical or current financial behaviors can be challenging, partly due to imbalanced data. The number of defaulters is relatively small compared to non-defaulters. Nonetheless, these cases are very important since they can cause huge losses if not predicted correctly. In this study, a comparison between four one-class classifiers and their combinations will be tackled, and several combining strategies will be explored. The one-class classification techniques isolation forest (IForest), support vector machine (SVM), Gaussian mixture model (GMM) and Parzen classifier (Parzen) and their hybrid models were applied on three different credit scoring datasets from uci machine learning repository; the Taiwanese dataset (30000 samples), the German dataset (1000 samples) and the Australian dataset (690 samples). The Australian dataset performed the best between the three datasets, especially when combining GMM and IForest classifiers, which gave an AUC result of 0.852 (sensitivity = 60.8%, specificity = 93.2%). Hybrid models enhanced the performance of one-class classifiers especially when including the strongest predictive classifier for the specific dataset. The study highlights the importance of hybrid models and their effect on improving the classification performances and reflects some interesting findings compared to past literature.