Dynamic Coupon Targeting Using Batch Deep Reinforcement Learning: An Application to Livestream Shopping.

The article focuses on developing and empirically validating a dynamic coupon targeting framework using batch deep reinforcement learning (BDRL) in a high-dimensional, high-frequency livestream shopping context. The BDRL approach models the dynamic coupon allocation problem as a Markov decision process, capturing consumers' intertemporal tradeoffs—particularly the reference price effect—without relying on parametric assumptions, thereby mitigating model bias. Tested on data from a major livestream shopping platform and through a large-scale field experiment, the BDRL policy increased customer lifetime value (CLV) by 63%, outperforming static targeting policies and a structural model-based dynamic policy. The framework effectively personalizes coupon discounts based on consumer heterogeneity and purchase dynamics, recommending strategies such as price skimming and adjusting discounts according to host attractiveness and consumer spending levels. The study highlights BDRL's potential for optimizing dynamic pricing and coupon strategies in e-commerce while acknowledging limitations related to forward-looking consumer behavior, unobserved heterogeneity, and data scope.