Rethinking Intermittent Demand Forecasting: Is Machine Learning the Future?

Demand forecasting is inherently challenging as it involves predicting irregular time series data that exhibit intermittent and lumpy patterns. This issue is especially pronounced when making forecasts at very granular levels, such as individual stores or warehouses, or for slow-moving items. Despite these difficulties, accurate forecasts are essential for effective inventory management and replenishment decisions. Traditionally, well-established statistical methods, such as Croston’s and its variants, have been used to tackle this problem. More recently, Machine Learning (ML) techniques have emerged as a potential alternative, though it is still debatable whether they offer a significant advantage. This paper contributes to the discussion by comparing the forecast performance of several ML approaches with traditional statistical methods using two extensive datasets of real demand data.