Dynamic Catalog Pape

Ontario and the 2002 Fall INFORMS Conference. We also gratefully acknowledge the contribution of the company that provided the data for this study and research assistance from Stephen Windsor. Deciding who should receive a mail-order catalog is amongst the most important decisions that mail-order catalogs must address. The current approach to the problem is invariably myopic: firms send catalogs to customers who they think are most likely to order from that catalog. In doing so the firms overlook the long-run implications of these decisions. For example, it may be profitable to mail to customers who are unlikely to order immediately if sending the current catalog increases the probability of a future order. We propose a model that allows firms to optimize mailing decisions by addressing the dynamic implications of their decisions. The model seeks near-optimal policies by taking advantage of recent advances in Approximate Dynamic Programming and Reinforcement Learning. In developing this model we identify and address generic methodological issues that arise in dynamic optimization when using a finite amount of data obtained under a specific historical policy. We apply the model to a large sample of historical data provided by a catalog firm. The historical data confirms the feasibility of the model, describes differences in the characteristics of the " optimal " and current mailing policies and provides predictions of the potential profit improvement.