Comparison of statistical and machine learning methods for daily SKU demand forecasting

Daily SKU demand forecasting is a challenging task as it usually involves predicting irregular series that are characterized by intermittency and erraticness. This is particularly true when forecasting at low cross-sectional levels, such as at a store or warehouse level, or dealing with slow-moving items. Yet, accurate forecasts are necessary for supporting inventory holding and replenishment decisions. This task is typically addressed by utilizing well-established statistical methods, such as the Croston’s method and its variants. More recently, Machine Learning (ML) methods have been proposed as an alternative to statistical ones, but their superiority remains under question. This paper sheds some light in that direction by comparing the forecasting performance of various ML methods, trained both in a series-by-series and a cross-learning fashion, to that of statistical methods using a large set of real daily SKU demand data. Our results indicate that some ML methods do provide better forecasts, both in terms of accuracy and bias. Cross-learning across multiple SKUs has also proven to be beneficial for some of the ML methods.