Contextual Offline Demand Learning and Pricing with Separable Models

This paper, inspired by a collaboration with a leading consumer electronics retailer in the Middle East, explores the challenge of demand learning and pricing using separable demand models. The data scarcity issue, characterized by limited price changes and low sales volumes, renders traditional models ineffective in deriving reasonable price elasticity. To address this issue, we advocate a separable model that leverages two submodels to distinctly capture the effects of price and contextual information. The separable structure enables us to invest special emphasis on the role of price and impose specific structural assumptions on the submodel for pricing effects, such as the monotone decreasing property. Theoretical analysis sheds light on the statistical complexity of demand learning with the separable structure, highlighting its capacity to reduce the necessary sample size to achieve a desired level of accuracy. We also introduce a computationally efficient iterative algorithm for deriving submodels from offline datasets, complete with convergence guarantees. In an empirical context, we demonstrate how our method can yield meaningful price elasticity estimations and revenue increase based on real sales data from the retailer.

This paper was accepted by J. George Shanthikumar, data science.

Funding: This work was supported by funding from the MIT Data Science Lab, Hong Kong Research Grants Council (RGC) General Research Fund [CityU11508223], and Hong Kong Research Grants Council (RGC) Early Career Scheme [CityU21505825].

Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2023.04026.