Using Bayesian Linear Models and Deep Neural Networks for Decomposing the Performance Effect of Promotion

Abstract

Price promotions can drive up short-term sales substantially. To establish whether and how a business can truly benefit from a price promotion bump, previous research has proposed multivariate linear regression models decomposing the performance effect of promotion into three constituent parts: cross-brand effects, cross-period effects, and category expansion effects. However, under missing data conditions, the original models fail to perform well, with a large part of constituent effects unexplainable. In this study, we propose a system of models that possess both explanatory and predictive power and can directly work on an imperfect dataset. Results show that the Bayesian linear regression models are able to conduct standard decomposition by demonstrating uncertainty and that our proposed deep neural networks drive predictive performance up to 42.14%.