Dynamic Policy Optimization for E-commerce Returns: A Reinforcement Learning Approach for SMEs with Limited Data

E-commerce returns represent a persistent challenge for small and medium-sized enterprises, with return rates averaging 17.6% across retail categories. I present a reinforcement learning framework that dynamically optimizes return policies for SMEs operating with limited transaction histories (10,000-100,000 records). My approach combines LASSO regression, Gradient Boosting, and customer segmentation with Q-learning to enable real-time policy adjustments. The framework incorporates pre-purchase interventions including augmented reality try-on features and AI-driven size recommendations. Testing on 100,000 German e-commerce transactions alongside deployment in an Indian marketplace showed 32% reduction in returns (from 24.7% to 16.8%), with prediction accuracy reaching 90.8%. The system achieved ROI between 109-736% while maintaining sub-100ms response times on standard cloud infrastructure. Through SHAP-based explainability, I demonstrate how SMEs can adopt sophisticated AI tools despite data constraints.