Working Paper: CEPR ID: DP16096
Authors: Adam Smith; Stephan Seiler; Ishant Aggarwal
Abstract: We study the profitability of personalized pricing policies in a setting with consumer-level panel data. To compare pricing policies, we propose an inverse probability weighted estimator of profits, discuss how to handle non-random price variation, and show how to apply it in a typical consumer packaged good market with supermarket scanner data. We generate pricing policies from Bayesian hierarchical choice models, regularized regressions, neural networks, and nonparametric classifiers using different sets of data inputs. We find that the performance of machine learning methods is highly varied, ranging from a 30.7% loss to a 14.9% gain relative to a blanket couponing strategy, whereas hierarchical models generate profit gains in the range of 13–16.7%. Across all models, information on consumers' purchase histories leads to large improvements in profits, while demographic information only has a small impact. We find that out-of-sample fit statistics are uncorrelated with profit estimates and provide poor guidance towards model selection.
Keywords: targeting; personalization; heterogeneity; choice models; machine learning
JEL Codes: C11; C33; C45; C52; D12; L11; L81
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| pricing strategies (D49) | profitability (L21) |
| personalized pricing methods (D49) | profitability (L21) |
| hierarchical models (C38) | profitability (L21) |
| purchase histories (D12) | profitability (L21) |
| demographic information (J10) | profitability (L21) |
| model fit statistics (C52) | profit estimates (C13) |
| hierarchical logit model (C35) | profitability (L21) |
| demographic variables (J10) | profitability in personalized pricing (D49) |