Machine Learning Learns Bayes

Abstract

Marketers today can choose between long-standing hierarchical Bayesian demand models and a fast-growing toolbox of machine learning methods to personalize prices. Yet guidance on their relative merits is scarce. We first introduce a semiparametric double/debiased machine-learning estimator of customer profit. Using this estimator, we then revisit coupon targeting in the IRI Marketing Dataset and compare eight pricing strategies. With the “off-the-shelf” feature set, Hierarchical Bayes attains the highest profit. However, when purchase history is passed as an encoder to the Machine Learners, the edge disappears. Our results stress the importance of processing information sets in solving Marketing problems.