When making marketing mix decisions, marketing managers of companies that offer a broad range of product categories, such as traditional offline and online retailers, mail-order companies, or financial service providers, often need to select one or a few focal categories out of all the possible ones offered. This interest is further fuelled by opportunities offered by the Internet or modern customer loyalty programs using smart card technologies, making it easier as well as cheaper for companies to implement micromarketing strategies.
These recent developments have lead to a shift in the managerial requirements of direct marketers: they want to find out which specific products or categories need to be featured in promotional activities customized for specific (groups of) customers. In this study, we present a decisionsupport tool that assists direct marketers in selecting subsets of promising categories from the large assortment they typically offer for inclusion in targeted promotions. The proposed analytical approach combines conventional wisdom of market basket analysis in a novel two-stage procedure (Boztug and Reutterer, 2008). In a first (exploratory) step, jointly purchased product categories across the entire assortment are identified by looking at pronounced cross-category interrelationships in the observed frequency patterns. Customers are next assigned to the identified shopping basket prototypes and we allow them to be members of multiple prototypes. This data-compression step is followed by a second (explanatory) step where the cross-category effects in response to marketing actions are modeled across the pre-selected categories. Our procedure takes both interdependencies in purchase behaviour across categories and customer heterogeneity with respect to cross-category effects in response to marketing actions into account.
For calibrating the model we obtained purchase transaction data of a major online grocery retailer for almost 4 year, resulting in a customer base of 17,312 households (purchased at least 3 times in the observation period). For the same retailer and time period, we also have detailed information on price and other important marketing-mix variables. A total number of 302,632 retail transactions with pick-any choices among an assortment of 121 categories are first subject to the data compression step. This first stage revealed 13 interesting and distinct prototypes which were subject to estimation of segmentspecific multivariate MNL models. Currently, we are about to empirically test the resulting recommendations derived from the above suggested two-stage approach vis-à-vis alternative approaches in a controlled field experiment conducted in cooperation with a major online grocery retailer. The experimental setup is projected to consist of a control group (no recommendations) and different experimental groups of which one group will receive recommendations derived by our suggested twostage approach and other groups will receive recommendations coming from other recommender systems that differ in their degree of intelligence. During the conference, we will present the underlying mechanism of our decision-support tool as well as show some preliminary results of its performance.