Bayesian assessment of predictors’ contributions to variation in the predictive performance of a logistic regression model

Abstract

ABSTRACT The logistic regression model is the algorithm most commonly applied in business analytics applications for classifying objects into binary categories among industrial users. This paper presents a Bayesian approach to assessing the contributions of predictors to the predictive performance of the classification model. Our proposed approach has two novel features that distinguish it from the usual approaches for such purpose. First, our approach ranks different predictors based on their contributions to variation in a model’s predictive performance, thus addressing the challenges of prediction risk and suggesting modelling strategy. Second, our approach can evaluate the contributions of every individual predictor each pair of two predictors. Hence, it can provide valuable information for managers on highly defined and detail-oriented business inquiries, complementary to the routine information conveyed by the usual methods for variable and feature selection purpose. We demonstrate the proposed approach using an example in credit risk management.