Dynamic Optimization of Catalog Distribution Decisions

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

The objective is to design optimal or near-optimal mailing policies for mail-order catalog companies. This is a complex problem because customer response functions are highly stochastic, reflecting in part the relative paucity of information that firms have about each customer. Moreover, the problem is a dynamic one. Customers' purchasing decisions are influenced not just by the firm's most recent mailing decision, but also by past mailing decisions. We will develop and test a model to optimize mailing decisions using the transaction and mailing histories of each customer. A moderately-sized discrete state space will be constructed to summarize these histories, leading to a tractable Markov Decision Process with two possible actions at each time step: mail or not mail. The proposed model will then be used to obtain near-optimal policies, by taking advantage of recent advances in Approximate Dynamic Programming and Reinforcement Learning. The results will be assessed and validated through a large-scale field-test. While developing this model, we will also identify and address generic methodological issues that arise in dynamic optimization when using a finite amount of data obtained under a specific historical policy. Catalog firms mail approximately 20 billion catalogs each year. Printing and mailing these catalogs is the second largest expense in the industry (behind the cost of the goods), representing approximately 20% of net sales. As a result, catalog managers view improving their policies for deciding who should receive catalogs as one of their highest priorities. However, because of the difficulty in identifying an optimal mailing policy, catalog firms typically only consider the immediate response when designing their policies. Initial tests suggest that by taking into account the long-run impact firms may be able to improve their profits by as much as 40%. While the proposed study will focus on catalog mailing decisions, these decisions are directly analogous to other types of retail decisions, including the distribution of retail coupons and other promotions. More generally, the project will advance our understanding of how to use historical data to improve the economic efficiency of the retail industry.