Retail Promotion Forecasting: A Comparison of Modern Approaches

Abstract

Promotions at retailers are an effective marketing instrument, driving customers to stores, but their demand is particularly challenging to forecast due to limited historical data. Previous studies have proposed and evaluated different promotion forecasting methods at product level, such as linear regression methods and random trees. However, there is a lack of unified overview of the performance of the different methods due to differences in modeling choices and evaluation conditions across the literature. This paper adds to the methods the class of emerging techniques, based on ensembles of decision trees, and provides a comprehensive comparison of different methods on data from a Danish discount grocery chain for forecasting chain-level daily product demand during promotions with a four-week horizon. The evaluation shows that ensembles of decision trees are more accurate than methods such as penalized linear regression and regression trees, and that the ensembles of decision trees benefit from pooling and feature engineering.