Deep Learning Models for Inventory Decisions: A Comparative Analysis

Over the past decade, a range of studies evaluated the benefits of considering machine learning methods and the power of auxiliary data to improve sales forecasting accuracy; however the analysis of how the forecasting predictions translate to lower cost inventory decisions is still in incipient stage. The focus of this contribution is to compare how different deep learning architectures leverage the potential of data features to achieve better demand estimation outputs and consequently contribute to a better optimization of single-period inventory decisions, also known as Newsvendor Problem. Additionally, we compare the performance of traditional model-based versus data-driven approaches for solving the Newsvendor Problem. We test the models in a real-world retail dataset and empirically show that recurrent architectures are the most successful in extracting the relevant features and processing large amounts of sequential data to achieve accurate forecasts that are used as input in the subsequent optimization step. Moreover, the data-driven inventory models outperformed in all the settings, providing the ordering decisions with the lowest mismatch costs.