DRIVER MODERATOR METHOD FOR RETAIL SALES PREDICTION

Abstract

We introduce a new method for stock keeping unit (SKU)-store level sales prediction in the presence of promotions to support order quantity and promotion planning decisions for retail managers. The method leverages the marketing literature to generate features, and data mining techniques to train a model that provides accurate sales predictions for existing and new SKUs, as well as consistent, actionable insights into category, store and promotion dynamics. The proposed "Driver Moderator" method uses basic SKU-store information and historical sales and promotion data to generate many features. It simultaneously selects few relevant features and estimates their parameters by using an L1-norm regularized epsilon insensitive regression that is formulated to pool information across SKUs and stores. Evaluations on two grocery store databases from Turkey and the USA show that out-of-sample predictions for existing and new SKUs are as good as, or more accurate than benchmark methods. Using the method's predictions for inventory decisions doubles the inventory turn ratio versus using individual regressions by lowering lost sales and inventory levels at the same time.