Disaggregated retail forecasting: A gradient boosting approach

Abstract

Demand forecast is a relevant topic for retailers to manage effectively inventories comprising a wide range of Stock Keeping Units (SKUs) at store level. While this disaggregated forecast problem is central to ensure profitability as it supports accurate inventory decisions, thus avoiding either out-of-stock events or overstocks and inventory losses, it is also very complex due to aspects such as the large number of stores and products of modern retailers, the complex marketing and promotional strategies that impact customer demand together with cross-product effects that are all difficult to model. In this study, we propose more effective methods to handle these aspects. More specifically, we employ XGBoost, a non-linear non-parametric ensemble-based model, as the central learning algorithm and a structural change correction method to account for sudden changes in consumer behavior caused by external factors. Our approach also encompasses data cleansing procedures to correct sales observations during out-of-stock days as well as discrepancies between logical and physical inventory counts. Based on real data from a public dataset of a large retailer, we show that our methods outperform the Base-Lift model, a widely used benchmark model for retail forecasting, yielding significant improvements in accuracy metrics together with reductions in stockouts and in stock on hand. The proposed approach has also a high degree of automation, an important requirement for modern retailers.