Forecasting Sales in Retail with XGBoost and Iterated Multi-step Ahead Method

Due to shopping trends change, retailers are prompt to optimize their business processes in order to provide more personalized, faster and smarter user experience, grow revenue and reduce business costs. Retail operational decisions include product allocation, product replenishment points and vehicle routes for inventory renewal. In all of these areas, a distinctive contribution lies in accurate estimation of product demand. This paper focuses on forecasting sales in retail. The product sales is modeled by using XGBoost algorithm and iterated multi-step ahead method on the horizon of 7 days. Model inputs include real historical sales data, seasonality and working/non-working day indicators. Model is tested with a real dataset of five chosen products provided by an industry partner. Results are compared to the baseline linear model and show improvement of over 21%.