A Machine Learning–Based Inventory Optimization Framework for Predictive Inventory Replenishment

Abstract

An effective inventory management system in for relief aid distribution requires an accurate demand forecasting and proactive replenishment strategies. However, existing approaches often address these demand predictions, safety stock computation, and restock decisions as separate processes that resulting in limited operational integration. This study proposes a machine learning–based inventory optimization framework for predictive inventory replenishment that addresses this gap by integrating forecasting and automated decision-making into a unified system. The framework utilizes a Gradient Boosting Regressor to predict continuous demand scores based on historical inventory data and engineered features. A priority classification mechanism is applied using the weighted demand index followed by priority-based safety factor mapping to differentiate inventory control policies. Safety stock is computed using forecast standard deviation, service factor, and lead time, while the reorder point (ROP) is calculated to determine replenishment thresholds. The automated restock decisions are generated by comparing inventory levels with the computed ROP. The framework was evaluated using a community dataset that comprises more than 76,000 records, resulting a small proportion of items (105 out of 16,200) were flagged for restocking and indicating an effective and targeted replenishment. The priority-based analysis further confirms that high-priority items are well-protected against demand variability, while low-priority items are managed efficiently to minimize excess inventory. Overall, the proposed framework demonstrates its capability to support data-driven and proactive inventory management that makes it suitable for relief operations and other dynamic supply chain environments.