Retail Demand Forecasting Using Spatial-Temporal Gradient Boosting Methods

Abstract

ABSTRACT With the significant growth of the e-commerce business, the retail industry is experiencing rapid developments, leading to the explosion of the number of stock-keeping units (SKUs). Therefore, it calls for forecasting algorithms to forecast a large number of product-level demands over a short forecasting horizon. We developed a novel machine learning algorithm—the spatial-temporal gradient boosting tree (ST-GBT)—for demand forecasting for the retail industry. By incorporating the cross-section and time-series information in the existing gradient-boosting decision tree algorithm, our new algorithm can accurately forecast tremendous SKUs in one process. Furthermore, we show potential factors related to the retail industry, while new factors, such as higher-order statistics and risk-free interest, are also proposed for demand forecasting tasks. The numerical experiment results based on a large e-commerce company’s historical transaction records support the comparative merits of the new algorithm with superior accuracy and automation ability.