Using reinforcement learning to optimize the acceptance threshold of a credit scoring model

This paper aims to study whether the reinforcement learning approach to optimizing the acceptance threshold of a credit score leads to higher profits for the lender compared to the state-of-the-art cost-sensitive optimization approach. We show that static methods, such as the latter do not ensure the optimality of the threshold leading to biased results and significant losses for the firm. We develop a dynamic reinforcement learning system that constantly adapts the threshold in response to live data feedback in order to maximize a company's profits. The developed algorithm is shown to outperform the traditional approach in terms of profits both in various simulated scenarios and using real data from an international consumer credit company. (C) 2019 Elsevier B.V. All rights reserved.