Retail demand forecasting using light gradient boosting machine framework

Abstract

Preparing product-level demand forecasts is crucial to the retail industry. Importantly, reliable in-ventory and replenishment decisions for retail products depend on accurate demand forecasts. This allows retailers to enable better pricing and timely promotion plans while leading to huge cost reductions [1]. Often, retail promotions create demand irregularities for products. Customers may change their buying behavior by purchasing more products for future consumption (stockpiling), thereby increasing sales in the promotional period. Then, for a brief time, sales may fall below normal levels before gradually returning to normal levels. The period with a dip in demand is known as the post-promotional period [2]. Thus, a retail promotion has three distinct periods: (1) normal, (2) promotional, and (3) post-promotional, each with its own set of demand fluctuations