Intermittent demand forecasting for spare parts with little historical information

Abstract

We propose a framework to address the demand forecast problem for spare parts with little or no historical demand information. This problem is relevant, e.g., in the context of forecasting the demand of parts in the initial procurement phase. We depart from the traditional forecasting approach by focusing on the distance between two demand profiles, rather than directly predicting the demand. That is, we build a model on the space induced by some distance metrics applied on attribute-wise comparisons of the spare parts, rather than directly working on the space induced by the original attributes. The three steps of the framework are: (i) data preparation, comparing parts along a set of attributes; (ii) model training, linking the attribute-wise distances and the distance between the demand profiles; (iii) neighborhood construction and forecasting, reconstructing the profile of the demand of the part with no historical demand. We tested the algorithm on a real-world instance with over 28 thousands spare parts from 131 Amazon sites, with the goal of reconstructing the demand of parts with no historical demand. Empirical evidence showcases the significant improvement brought about by the proposed approach, compared with a pool of alternative models.

Keywords:

• Spare parts
• forecasting
• machine learning
• distance metrics

Disclosure statement

The authors report there are no competing interests to declare.