An ensemble credit scoring model based on logistic regression with heterogeneous balancing and weighting effects

Abstract

The logistic regression model is widely used in credit scoring practice due to its strong interpretability of results, but its recognition performance for default samples which are minority in real-world imbalanced data sets need to be improved. This paper designs a novel ensemble model based on logistic regression as the logistic-BWE model. It first carries out data preprocessing, then applying sample balancing algorithm to generate several training sub data sets with different imbalance ratios and constructing sub models respectively, finally according to the performance of each sub model in the validation stage, the weight of predicted results for different class of each sub model is dynamically calculated. The empirical results indicate that compared with ten representative credit scoring models on six public data sets, the logistic-BWE model has the strongest ability to recognize default samples, and has the best generalization ability on most data sets while maintaining the interpretability. Further tests demonstrate that the performance superiority of the logistic-BWE model is statistically significant, and it also has excellent robustness when it contains a sufficient number of sub models.