A0999
Title: Reinforcement learning in credit risk management
Authors: Jorge C-Rella - University of A Coruna (Spain) [presenting]
David Martinez Rego - DataSpartan (United Kingdom)
Juan Vilar Fernandez - Universidade da Coruna (Spain)
Abstract: Credit risk problems are dynamic because customer behavior is not stable, and they are cost-sensitive in the sense that a decision's impact depends on the loan amount. In addressing these challenges, online learning algorithms serve as valuable tools, adapting in real-time as new data becomes available. However, relying solely on approved transactional data introduces potential unfair biases and opportunity costs. Within reinforcement learning, bandit algorithms offer a solution by effectively balancing the trade-off between exploiting current models and exploring actions with limited information to improve predictions. Novel dynamic learning strategies are presented, which extend online learning and bandit algorithms to cost-sensitive classification. Empirical evaluations conducted on benchmark datasets and through extensive simulation studies corroborate the effectiveness and efficiency of the proposed methodologies.
