Enhancing intermittent spare part demand forecasting : a novel ensemble approach with focal loss and smote

Background: Accurate inventory management of intermittent spare parts requires precise demand forecasting. The sporadic and irregular nature of demand, characterized by long intervals between occurrences, results in a significant data imbalance, where demand events are vastly outnumbered by zero-demand periods. This challenge has been largely overlooked in forecasting research for intermittent spare parts. Methods: The proposed model incorporates the Synthetic Minority Oversampling Technique (SMOTE) to balance the dataset and uses focal loss to enhance the sensitivity of deep learning models to rare demand events. The approach was empirically validated by comparing the model's Mean Squared Error (MSE) performance and Area Under the Curve (AUC). Results: The ensemble model achieved a 47% reduction in MSE and a 32% increase in AUC, demonstrating substantial improvements in forecasting accuracy. Conclusions: The findings highlight the effectiveness of the proposed method in addressing data imbalance and improving the prediction of intermittent spare part demand, providing a valuable tool for inventory management.