Intermittent demand forecasting with transformer neural networks

Intermittent demand forecasting is an important yet challenging task in many organizations. While prior research has been focused on traditional methods such as Croston’s method and its variants, limited research has been conducted using advanced machine learning or deep learning methods. In this study, we introduce Transformer, a recently developed deep learning approach, to forecast intermittent demand. Its effectiveness is empirically tested with a dataset of 925 intermittent demand items from an airline spare parts provider and compared with that of two traditional methods such as Croston’s and the Syntetos–Boylan approximation as well as several popular neural network architectures including feedforward neural networks, recurrent neural networks, and long short-term memory. Our results based on six different forecasting performance measures show that Transformer performs very well against other methods in a variety of settings. We also examine how data sparsity impacts model performance and find that different models perform similarly when sparsity is low. Although the performance of all models generally gets worse as the sparsity level increases, the advantage of Transformer over other models increases with sparsity.