Intermittent demand forecasting remains a critical challenge in time series analysis due to its sparse and unpredictable nature. This paper introduces a two-step framework combining Convolutional Neural Networks and XGBoost for forecasting. Three model architectures are evaluated: a dual CNN model, a hybrid CNN-XGBoost model, and a fully XGBoost-based model. Using a real-world dataset of 4,406 raw materials from a tailor-made production company, the models are assessed via multiple metrics, including MAE, RMSE, R², and Zero Prediction Accuracy. Results indicate that while the XGBoost model yields lower raw errors, the hybrid model achieves superior explanatory power. This research demonstrates the advantages of integrating deep learning and ensemble methods for more robust Intermittent demand forecasting.
Mirshahi et al. (Thu,) studied this question.