When to Target Customers? Retention Management using Dynamic Off-Policy Policy Learning

We examine how to learn personalized customer retention strategies when customers' intentions to purchase evolve over time. Working with a Japanese online platform, we first implement a large-scale randomized experiment, in which coupons are randomly sent to first-time buyers at different times. The experimental data allows us to estimate the personalized dynamic retention policy using the off-policy policy learning methods. We extend the existing methods by allowing non-Markovian strategies and by considering inter-temporal budget constraints. Our results show that the optimal dynamic policy is more cost-effective than baseline policies. We also test the optimal policy online to confirm its performance