Machine Learning for Inventory Management: Analyzing Two Concepts to Get From Data to Decisions

We analyze two fundamentally different concepts to considering data for planning decisions using the example of a newsvendor problem in which observable features drive variations in demand. Our work contributes to the extant literature in two ways. First, we develop a novel joint estimation-optimization (JEO) method that adapts the random forest machine learning algorithm to integrate the two steps of traditional separated estimation and optimization (SEO) methods: estimating a model to forecast demand and, given the uncertainty of the forecasting model, determining a safety buffer. Second, we provide an analysis of the factors that drive difference in the performance of the corresponding SEO and JEO implementations. We provide the analytical and empirical results of two studies, one in a controlled simulation setting and one on a real-world data set, for our performance evaluations. We find that JEO approaches can lead to significantly better results than their SEO counterparts can when feature-dependent uncertainty is present and when the cost structure of overage and underage costs is asymmetric. However, in the examined practical settings the magnitude of these performance differences is limited because of the overlay of opposing effects that entail the properties of the remaining uncertainty and the cost structure.