Using Deep Learning to Improve Inventory Record Accuracy: Concept and Application

Abstract

Maintaining accurate inventory records has been, and remains, a central problem for managing retail operations. Discrepancies between the physical and recorded stock lead to poor reordering decisions, the over- or understocking of products that increase waste or lost sales, respectively. In this study, we revisit this classic inventory management problem by investigating whether novel machine learning algorithms provide an improvement over established methods. Specifically, we explore the application of deep learning as a method to identify and correct impactful inventory record errors – those that may affect future reordering decisions – by leveraging product level, store level and inventory quality data. Our results show that a deep learning-based decision model outperforms traditional stock auditing strategies. We validate our model by running a field experiment on 450 large stores of a leading grocery retailer. The results demonstrate that a deep learning-based decision model can provide a significant improvement over traditional corrective strategies by refining the representation of the inventory record inaccuracy problem. More generally, our study highlights how generic algorithms, like deep learning, can be applied successfully to improve performance over existing specific methods, yet also, we identify the wider challenges that emerge when deploying those algorithms.