An Uncertainty-Aware Reinforcement Learning Framework for Joint Demand Forecasting and Inventory Optimization

Baqiang Chen^*Baqiang Chen^*

• Sichuan Academy of Social Sciences, Chengdu, China

The optimization of inventory systems is a cornerstone of operational efficiency, yet it is persistently challenged by demand uncertainty. Traditional approaches that decouple demand forecasting from inventory control often propagate forecast errors, leading to suboptimal and brittle policies. In this paper, we propose a novel, integrated framework that synergizes probabilistic deep learning for forecasting with uncertainty-aware reinforcement learning (RL) for dynamic inventory optimization. Our primary innovations are threefold: (1) a probabilistic forecasting model (a Bayesian LSTM) that quantifies predictive uncertainty, providing not just a point forecast but a full predictive distribution; (2) an uncertainty-augmented state representation for the RL agent, which explicitly includes the forecasted demand's mean and variance, enabling it to learn risk-sensitive behaviors; and (3) a dynamic risk-adjusted reward function that penalizes the agent more severely for failures under low uncertainty, encouraging it to be robust when predictions are confident and cautiously opportunistic when they are not. We employ a Deep Deterministic Policy Gradient (DDPG) agent that learns a continuous, state-dependent order-up-to-level policy. Evaluated on real-world retail data and a series of challenging synthetic scenarios, our framework demonstrates marked superiority over traditional and state-of-the-art baselines. It achieves a 28-40% reduction in total inventory-related costs and improves the service level by 5-10 percentage points, while simultaneously reducing order volatility. The learned policies are shown to be both interpretable and economically rational. This work provides a new paradigm for building intelligent, adaptive, and risk-aware decision-making systems for supply chain management, showcasing how AI can effectively manage and leverage uncertainty to drive significant economic value.