Seasonal Marketing Campaigns : Rethinking Exploration and Exploitation with Infrequent Large Batches
Policies that guide how firms trade off exploring to improve a current targeting policy, versus exploiting to harvest profits from the current policy, typically focus on settings in which individual or small batches of customers arrive frequently. However, when demand is seasonal, marketing campaigns often occur annually, with retailers using data from last year to train this year’s policy. This has two implications for how firms resolve the exploration exploitation trade off. First, many customers in each seasonal batch introduces an information externality: the incremental information contributed by a focal customer depends upon the assignment decisions for other customers in the batch. Second, there is a long time between batches. This means that periods beyond the next batch are in the distant future, and so it becomes reasonable to approximate the problem by looking just one batch ahead. We investigate how to optimally rebalance exploration and exploitation in these settings. The algorithm we propose uses Gaussian Processes to estimate the value of additional exploration, while accounting for the information externality between customers in the same batch. We validate our findings using data from a field experiment