Default Prediction Framework With Optimal Feature Set and Matching Ratio

Abstract

We propose a default prediction framework that incorporates imbalance handling and feature selection. For imbalance handling, we determine the optimal ratio of non-default to default firms by minimizing the Type-II error of the majority voting deep fully connected network (MV-DFCN) model. For feature selection, we design a two-stage process that first eliminates highly correlated and redundant features, and then refines the feature set using backward selection. Experimental results show that the DFCN model within the proposed framework outperforms baseline models in terms of G-Mean and AUC and achieves the lowest Type-II error rate. Furthermore, the framework outperforms eight baseline combinations of imbalance handling and feature selection strategies. Additionally, SHAP values are used to assess feature contributions, and nine features with statistically significant impacts are identified.