Forecasting Intermittent Demand for Aircraft Spare Parts using Machine Learning

Abstract

Intermittent demand poses a complex challenge for military aviation logistics due to long periods of zero consumption, followed by sudden peaks of demand, which makes traditional forecasting methods unreliable. This thesis develops and evaluates machine-learning approaches for forecasting intermittent aircraft spare parts demand within the Brazilian Air Force (FAB), with the goal of improving prediction accuracy and, consequently, enhancing readiness levels and reducing stockouts. The study gathers and preprocesses 10 years of historical spare-parts demand data from the T-27 (EMB-312) TUCANO aircraft fleets and applies different forecasting models, including Moving Average, Simple Exponential Smoothing, and Croston’s method as traditional time-series baselines, and gradient boosted decision trees (XGBoost), and artificial neural networks as machine learning models. Accuracy is assessed using MASE as the primary metric, complemented by error distributions. Results demonstrate that machine learning models, particularly XGBoost combined with engineered features, achieve significant gains over classical methods in forecasting accuracy. The findings provide a replicable framework for modernizing FAB’s spare parts planning process and highlight opportunities for broader adoption of advanced analytics in defense logistics.