Reinforcement Learning for Dynamic Credit Scoring Systems: Design, Implement, Evaluate

Thesis Abstract

This study investigates the use of reinforcement learning (RL) to design, implement, and evaluate dynamic credit scoring systems capable of adapting to evolving borrower behavior and macroeconomic conditions. The problem addressed is the static nature of traditional credit scoring models, which often fail to respond promptly to regime shifts, leading to suboptimal risk assessment and elevated default rates. The aim is to develop an RL-driven credit scoring framework that optimizes long-term portfolio performance while maintaining acceptable profitability and fairness. Specific objectives include (1) developing an RL agent that dynamically updates credit decision thresholds in response to changing risk signals; (2) integrating feature representations from historical borrower data with state-of-the-art RL algorithms (e.g., proximal policy optimization and deep Q-learning) to produce calibrated approval likelihoods; (3) evaluating the framework against conventional logistic regression and gradient boosting models in terms of predictive accuracy, default rate, and economic value added; (4) examining the robustness of the approach under covariate shift and varying macroeconomic scenarios; and (5) assessing implementation feasibility in real-time banking environments with policy constraints and explainability requirements. The methodology adopts an experimental design grounded in a simulated but realistic banking environment, augmented by a retrospective real-world data study. The population comprises consumer loan applicants from a mid-size commercial bank, with a dataset of 2.5 million credit applications spanning five years (including default outcomes, credit limits, payment histories, and macroeconomic indicators). A stratified sampling approach yields a representative sample of 1,000,000 applications for model development and 500,000 for out-of-sample testing, ensuring coverage across credit segments and economic cycles. Data collection relies on bank-generated transactional and credit bureau records, supplemented by publicly available macro indicators (GDP growth, unemployment, inflation). Instruments include feature extraction pipelines for borrower risk signals (e.g., utilization ratios, payment delinquency streaks) and environment state variables (interest rate shifts, unemployment spikes). Validity and reliability are ensured through data cleaning protocols, cross-validation, and backtesting against holdout periods. The RL framework leverages a Markov decision process where the agent selects credit actions (approve/deny, credit limit adjustments, pricing) given state representations comprising borrower features and macroeconomic indicators. Policy optimization employs a hybrid approach a neural network policy trained via proximal policy optimization (PPO) to balance risk-reward trade-offs, and a value network for baseline estimation. The agent undergoes pretraining on historical data using batch RL, followed by online fine-tuning within a controlled simulation environment that models borrower responses, delinquency dynamics, and portfolio-level constraints (capital adequacy, regulatory limits). Baseline comparisons include logistic regression, gradient boosting (XGBoost), and a static credit scoring model, with performance evaluated on predictive accuracy (AUC), calibration (Brier score), and economic value added (EVA). Hypothesis testing employs paired t-tests and bootstrap confidence intervals for key metrics, while sensitivity analyses examine covariate shift and policy constraint variations. Explainability is addressed via SHAP value decomposition and counterfactual analysis to translate RL decisions into human-understandable risk factors. Key expected findings include (i) the RL-based dynamic scoring system achieving superior cumulative profitability and lower expected portfolio loss than static models, (ii) improved responsiveness to macroeconomic shocks evidenced by reduced deterioration in approval quality during downturns, and (iii) acceptable calibration with transparent risk drivers enabling regulatory compliance. The study anticipates that PPO-based policies will yield smoother approval boundaries and more resilient pricing strategies, while maintaining fairness across credit segments through constraint-aware optimization. Contributions to knowledge encompass (a) a novel integration of reinforcement learning with credit risk decision-making under regulatory and business constraints; (b) a rigorous evaluation protocol for dynamic scoring under regime shifts and real-world noise; and (c) practical insights into deployment considerations, including model governance, explainability, and ethical risk management. The main conclusion is that reinforcement learning can operationalize dynamic credit scoring with measurable gains in portfolio performance and risk control, given careful design, robust validation, and transparent policy constraints. Recommendations emphasize staged deployment with ongoing monitoring, incorporation of fairness metrics, continuous data governance, and extension to other lending products and cross-border portfolios.