In this study, the authors conduct an experiment to simultaneously address two main issues that hinder credit default prediction: missing variables and imbalanced data. The objective is to determine which combination of solutions performs best under specific scenarios. For the first issue, data are imputed using various algorithms, including Random Forest, Classification and Regression Tree, and Mode-Median values, following the implementation of missing data generation through MNAR, MAR, and MCAR mechanisms. Regarding the second issue, to enhance the predictive power of default versus non-default classification in the context of imbalanced data, methods such as Down-sampling, Synthetic Minority Over-sampling Technique, and Cost Sensitive Learning are integrated into the final classifiers, including K-Nearest Neighbors, Random Forest, Support Vector Machine, and traditional Logit regression. The performance of different combinations of these treatments is evaluated and compared using the f-measure, which has been shown to provide more accurate comparison results when dealing with imbalanced data. The findings indicate that the imbalanced class problem significantly impacts model predictive performance. Among the three imputation methods, MNAR emerges as the most effective.
Bin Li (Fri,) studied this question.