Credit Risk Scoring with Bayesian Network Models

This paper proposes a Bayesian network model to address censoring, class imbalance and real-time implementation issues in credit risk scoring. It shows that the Bayesian network model performs well against competing models (logistic regression model and neural network model) along several dimensions such as accuracy, sensitivity, precision and the receiver characteristic curve. Better performance of the Bayesian network model is particularly salient with class imbalance, higher dimensions and a rejection sample. Furthermore, the Bayesian network model can be scaled efficiently when implemented onto a larger dataset, thus making it amenable for real-time implementation.