A machine learning approach to inventory stockout prediction

Abstract

The retail industry continues to experience frequent stockouts, driven by the rise of e-commerce and disruptive events such as the COVID-19 pandemic, which have significantly impacted both profitability and supply chain stability. As a result, developing effective models for stockout prediction has become increasingly critical for enhancing the efficiency and resilience of retail operations. The growing availability of data, challenges posed by data imbalance, and high demand uncertainty underscore the need to transition from traditional forecasting models to more intelligent, data-driven approaches that integrate multiple relevant features alongside sales data. In this study, we utilise a large dataset from a retailer comprising over 1.6 million stock keeping units (SKUs) to develop an analytical model based on classical machine learning algorithms aimed at improving stockout prediction accuracy. Our results demonstrate that the proposed approach performs well in handling large-scale, imbalanced data and significantly enhances predictive performance. Feature importance analysis reveals that current inventory levels, short-term demand forecasts (three months), and recent sales data are the most influential factors in predicting stockouts. Furthermore, the findings suggest that recent demand forecasts and sales data have greater predictive power than longer-term projections (six and nine months), highlighting the importance of near-term indicators in inventory stockout prediction accuracy. To the best of our knowledge, these insights provide valuable contributions to understanding stockout dynamics and improving inventory management strategies within the retail sector.