XGBoost modeling for sparse spare-parts demand forecasting

Abstract

Spare parts demand in many industrial systems is inherently sparse and intermittent. In practice, long periods of zero usage are common, even though inventory must still be maintained to ensure operational reliability. This situation increases holding costs and the risk of obsolescence, while also limiting the effectiveness of conventional forecasting techniques. This study demonstrates that a global XGBoost model trained across multiple spare-part items significantly outperforms item-specific models under sparse demand conditions. Using six years of historical spare-parts usage and procurement data from the energy sector, this study compares global and single-item extreme gradient boosting (XGBoost) modeling strategies. Forecast accuracy is evaluated using mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and median absolute error (MdAE), which is particularly suitable for zero-inflated demand patterns. The results consistently show that the global XGBoost model achieves lower errors across all metrics. In particular, the global model attains a markedly lower MdAE (0.00018), indicating greater robustness when demand is irregular and intermittent.