Demand forecasting for military Maintenance, Repair, and Overhaul (MRO) consumables is uniquely challenging: Parts are expensive, lead times are long, and a stockout can ground aircraft rather than result in a lost sale. This project partners with the Defense Logistics Agency (DLA) to improve forecast accuracy for U.S. Naval Aviation MRO consumables. The proposed framework clusters Stock Keeping Units (SKUs) by volume, volatility, and intermittency using K-means, then evaluates 25 forecasting models for each SKU, ranging from Naive baselines to deep learning architectures, and selects the winners based on Mean Absolute Error (MAE). Accuracy improvements over the Naïve Mean baseline range from 16% to 59%, depending on the cluster. Probabilistic forecasts from the top three models per cluster feed directly into a base-stock inventory policy, enabling planners to set order-up-to levels to meet a target service level. The study demonstrates that the cluster-then-forecast paradigm, previously validated in commercial settings, extends effectively to defense logistics and provides DLA with an operational tool to reduce both stockouts and excess inventory across the naval aviation supply chain.
