Authors: Sakthi Sankara Balaji Sathyamurthy
Abstract: Wholesale banking credit decisions involve large and interconnected exposures, evolving borrower conditions, collateral movements, covenant compliance, and changing market environments. Conventional credit-risk models are commonly designed for one-time prediction tasks and may not adequately support sequential decisions such as adjusting exposure limits, revising pricing, requesting additional collateral, or escalating cases for review. This study proposes an explainable reinforcement learning framework for real-time credit-risk decisioning in wholesale banking ecosystems. The framework represents credit management as a sequential decision problem in which an agent observes dynamic borrower, facility, portfolio, and market conditions before recommending appropriate risk actions. It integrates offline reinforcement learning with policy constraints to reduce the risk of unsafe or non-compliant recommendations in regulated banking environments. The proposed model incorporates an explainability layer that provides feature-based explanations, reward decomposition, policy rationale, and counterfactual action analysis. These outputs are intended to clarify why a specific credit action is recommended, which risk factors influenced the decision, how the recommendation aligns with risk appetite, and what changes in borrower conditions could lead to an alternative outcome. The study adopts a design science and empirical evaluation approach using historical wholesale credit data, with performance assessed against conventional statistical models, machine- learning models, rule-based decision engines, and non-explainable reinforcement-learning approaches. Evaluation criteria include credit-risk performance, expected-loss control, risk-adjusted return, decision latency, policy compliance, robustness, and explanation quality. The study contributes a structured framework for applying explainable reinforcement learning to high-impact credit decisions. It also provides guidance for controlled deployment through offline training, human oversight, model validation, audit logging, and continuous monitoring. The proposed approach may strengthen the timeliness, consistency, transparency, and governance of credit-risk decisioning across wholesale banking operations.
