On the Relevance of Demand Pattern Categorization

Abstract

The application of transfer learning to predict sales demand is an emerging topic that has been attracting more and more attention recently. However, the selection of data to be used for the learning process is not trivial. Data resources are usually scarce and often anonymized to a certain extent, so their usability for successful training is not guaranteed. One solution is to use already developed categorization schemes that group time series based on certain calculated parameters, but the derived categories do not necessarily capture the process of time series formation. This research addresses the question of whether categorization schemes are beneficial for transfer learning approaches by conducting an experiment in which Syntetos’, Boylan’s and Croston’s categorization scheme is used in combination with two deep learning architectures for the transfer learning process. The results show that similar patterns are indeed beneficial for prediction, but that models using all available data perform quite similarly.