Abstract

A novel framework for profit-based credit scoring is proposed in this work. The approach is based on robust optimization, which is designed for dealing with uncertainty in the data, and therefore is effective at classifying new samples that follow a slightly different distribution in relation to the original dataset used to create the model. Instead of minimizing a loss function based on statistical measures, the proposed method maximizes the profit of the credit scoring model, balancing the benefits and losses of granting credit with the variable acquisition costs. The reduction of these is performed using feature selection techniques embedded in the learning process. The robust approach results in four second order cone programming formulations, which can be solved efficiently using interior point algorithms. Experiments on two credit scoring datasets demonstrate the virtues of our approach in terms of its predictive performance, and the managerial insights that can be gained from it. (C) 2019 Elsevier Inc. All rights reserved.