Gradient Boosting-Based Demand Variability Estimation for Improved Safety Stock Calculation in Multi-Echelon Aerospace Spare Parts Inventory

Accurate safety stock calculation in multi-echelon aerospace spare parts networks remains a persistent challenge due to intermittent demand patterns and stochastic lead time variability. Traditional statistical approaches to demand variance estimation often fail to capture the complex, non-linear dependencies inherent in aerospace aftermarket environments. This paper investigates the application of gradient-boosting-based machine learning techniques---specifically LightGBM and XGBoost---to improve the estimation of demand variability and lead time variance as direct inputs to safety stock formulas in a three-echelon inventory network. A comparative evaluation is conducted across five estimation methods using a dataset of 2,847 aerospace spare part SKUs spanning 60 months of transactional records. The experimental results indicate that LightGBM reduces the error in estimating demand standard deviation by 18.6% relative to exponential smoothing, translating to a 9.3% reduction in total inventory holding cost while meeting echelon-specific cycle service level (CSL) targets across the network. A component criticality-based stratification analysis further reveals that gradient boosting methods yield the most pronounced improvements for intermittent-demand, high-criticality parts. The findings provide empirical evidence supporting the integration of machine-learning-enhanced variance estimation into existing multi-echelon safety-stock optimization frameworks for aerospace logistics applications.