Constructing learning models from data : the dynamic catalog mailing problem

Abstract

The catalog industry is a large and important industry in the US economy. One of the most important and challenging business decisions in the industry is to decide who should receive catalogs, due to the significant mailing cost and the low response rate. The problem is a dynamic one - when a customer is ready to purchase, s/he may order from a previous catalog if s/he does not have the most recent one. In this sense, customers' purchasing behavior depends not only on the firm's most recent mailing decision, but also on prior mailing decisions. From the firm's perspective, in order to maximize its long-term profit it should make a series of optimal mailing decisions to each customer over time. Contrary to the traditional myopic catalog mailing decision process that is generally implemented in the catalog industry, we propose a model that allows firms to design optimal dynamic mailing policies using their own business data. We constructed the model from a large data set provided by a catalog mailing company. The computational results from the historical data show great potential profit improvement. This application differs from many other applications of (approximate) dynamic programming in that an underlying Markov model is not a priori available, nor can it be derived in a principled manner. Instead, it has to be estimated or "learned" from available data. The thesis furthers the discussion on issues related to constructing learning models from data. More specifically, we discuss the so called "endogeneity problem" and the effects of inaccuracy in model parameter estimation. The fact that the model parameter estimation depends on data collected according to a specific policy introduces an endogeneity problem. As a result, the derived optimal policy depends on the original policy used to collect the data.

(cont.) In the thesis we discuss a specific endogeneity problem, "attribution error." We also investigate whether online learning can solve this problem. More specifically, we discuss the existence of fixed point policies for potential on-line learning algorithms. Imprecision in model parameter estimation also creates the potential for bias. We illustrate this problem and offer a method for detecting it. Finally, we report preliminary results from a large scale field test that tests the effectiveness of the proposed approach in a real business decision setting.