Identifying drivers of forecasting model performance for highly intermittent SKU demand

Abstract

In the area of demand forecasting, stock-keeping unit (SKU) demand forecasting for inventory management presents a unique challenge to the forecaster because of its high levels of intermittency (frequent zero-demand periods). While the literature on regular demand forecasting has made great progress in categorizing and concretizing the drivers of forecasting accuracy, encompassing research on intermittent demand is still scarce. This work identifies and assesses such drivers for intermittent demand by integrating the most recent related findings for regular demand with the few findings on high-performing models in intermittent demand. It assesses the benefit of machine learning (ML) models, statistical models developed specifically for intermittent demand, and the performance drivers temporal aggregation (TA), external variables, and cross-learning (CL), a model modification in which the model is not trained on a single time series but in vector-model fashion on multiple time series simultaneously, on a data set of highly intermittent products from the automotive spare parts industry. Results confirm the superiority of ML models such as support vector regression and XGboost and the benefits of TA. They also make an argument for the existence of statistical models for intermittent demand in comparison to conventional statistical models. Findings for external variables and CL are inconsistent and show that their benefit largely depends on the selected ML model. This research also provides useful practical recommendations for the selection of the adequate forecasting model. Interestingly, it finds that a product-based model selection approach does not improve overall forecasting performance as much as one would expect in comparison to a more generic selection approach.