Loyalty programs for convenience stores generate consumer shopping histories that are both large in size and sparse in content. Analyzing such data with traditional basket models is computationally difficult since most models are not scalable to a large set of categories. However, analyzing large data with traditional models has important advantages: the models capture consumer (shopping) behaviors that assist managers in making strategic decisions. In this thesis, we develop two studies to analyze this large and sparse convenience store shopping data.
In the first study, we bridge the gap between traditional basket model analysis and the challenges of large shopping data by developing a retail market basket modeling system that captures essential elements of consumer shopping behavior in a computationally attractive manner. An application of the model to convenience store basket data yields excellent results. The main outputs of the model (segmentation structure, cross-category dependence, price elasticities) align well with managerial intuition. Moreover, the model provides excellent forecasts to a holdout sample of consumers. Using the model, we examine the revenue impact of a change in promotion policy.
In the second study, we add spatial extensions to the previous model to solve a more complex problem: retail location analysis. We develop a spatial basket model to analyze the spatial pattern of consumer heterogeneity across stores, and show how to use this model to predict the demand of a new store (without any data of consumer purchase history). The main outputs of the extended model also align well with managerial intuition. Additionally, the model provides excellent forecasts to the demand of the hold-out store.