Bridging the Gap: Enhancing Reinforcement Learning-Based Stock Trading with Explainable AI

Profit alone is insufficient in financial trading; Algorithmic Trading systems must also demonstrate risk-awareness, transparency, generalisation. Existing Reinforcement Learning (RL) approaches optimise for return, often sacrificing stability and explainability, limiting practical use. We propose a sector-aware RL trading system with three innovations: (i) Universal Base (UB) initialisation, warm-starting agents with sector level facts (i.e., basic movements, trends common to stocks in a sector) before specialising to individual assets; (ii) an equal-weight (EqualW) allocation ensuring fair benchmarking without mega-cap dominance; and (iii) SHAP explainability for auditable decisions. Tested first on a small subset, and then the NASDAQ-100, UB EqualW portfolios sustained favourable risk-adjusted outcomes. SHAP revealed that UB relied on balanced technical, sentiment, and macroeconomic inputs, unlike non-UB models that overfit noisy signals. Overall, this framework delivers interpretable, risk-aware RL trading policies that generalise across regimes, combining robust returns with the transparency required in financial markets.