Thursday, January 20, 2022
We examine the proﬁtability of personalized pricing policies that are derived using diﬀerent speciﬁcations of demand in a typical retail setting with consumer-level panel data. We generate pricing policies from a variety of models, including Bayesian hierarchical choice models, regularized regressions, and classiﬁcation trees using diﬀerent sets of data inputs. To compare pricing policies, we employ an inverse probability weighted estimator of proﬁts that explicitly takes into account non-random price variation and the panel nature of the data. We ﬁnd that the performance of machine learning models is highly varied, ranging from a 21% loss to a 17% gain relative to a blanket couponing strategy, and a standard Bayesian hierarchical logit model achieves a 17.5% gain. Across all models purchase histories lead to large improvements in proﬁts, but demographic information only has a small impact. We show that out-of-sample hit probabilities, a standard measure of model performance, are uncorrelated with our proﬁt estimator and provide poor guidance towards model selection.