Prescriptive Analytics for Inventory Management: A Comparison of New Approaches

Abstract

We analyze the performance drivers for data-driven inventory management in a Newsvendor setting with nonstationary demand. For this, we study two novel approaches which are based on machine learning techniques (linear quantile regression and tree-based regression, respectively) and which use historical demand observations and auxiliary data to prescribe optimal inventory quantities. We identify three major performance drivers, that are non-linearity, heteroscedasticity and usability. We evaluate both models both in an extensive simulation experiment where we control different properties of the feature-demand relationship as well as on a complex real-world data set from a restaurant chain. From these experiments we conclude that in situations in which the structure of the feature-demand relationships is not known to be predominantly linear - which can be assumed to be the typical case in practice - we recommend the much more robust tree-based approach. Furthermore, the tree-based also provides comparatively better results for feature-dependent noise, i.e., heteroscedasticity. We also find that in the real-world setting, the tree-based approach performs better with very small data-sets due to the better built-in feature selection which is important in terms of usability in practice.

Keywords: Prescriptive Analytics, Inventory Management, SEP Newsvendor Model, Machine Learning, Tree-Based Regression, Linear Quantile Regression