Credit scoring in Banking: Reinforcement Learning Strategies for Profit Performance and Default Risk Management

Credit scoring is essential in financial risk management, yet traditional approaches often rely on rule-based or statistical models that do not adapt to dynamic behavior of the borrower. Recent reinforcement learning (RL) models offer improvements, but primarily focus on single-objective profit maximization. However, credit scoring is inherently a complex multi-objective problem, and most existing models depend on detailed customer-level modeling, which is costly and not easily scalable. To address these limitations, this study proposes a multi-objective reinforcement learning framework for optimizing credit scoring decisions by jointly maximizing expected profit, minimizing credit loss, and balancing approval rates. The model is trained entirely on historical data, eliminating the need for real-time customer engagement. It also incorporates non-interest income components such as merchant service charges to reflect actual profitability better and introduces a proactive risk mitigation strategy by deactivating remaining credit availability for high-risk borrowers before expected default. Experiments on multiple datasets show that the proposed Double Q-Learning model outperforms traditional baselines, offering a scalable, data-driven solution for adaptive and profit-aware credit scoring.