Two New Prediction-Driven Approaches to Discrete Choice Prediction

The ability to predict consumer choices is essential in understanding the demand structure of products and services. Typical discrete choice models that are targeted at providing an understanding of the behavioral process leading to choice outcomes are developed around two main assumptions: the existence of a utility function that represents the preferences over a choice set and the relatively simple and interpretable functional form for the utility function with respect to attributes of alternatives and decision makers. These assumptions lead to models that can be easily interpreted to provide insights into the effects of individual variables, such as price and promotion, on consumer choices. However, these restrictive assumptions might impede the ability of such theory-driven models to deliver accurate predictions and forecasts. In this article, we develop novel approaches targeted at providing more accurate choice predictions. Specifically, we propose two prediction-driven approaches: pairwise preference learning using classification techniques and ranking function learning using evolutionary computation. We compare our proposed approaches with a multiclass classification approach, as well as a standard discrete choice model. Our empirical results show that the proposed approaches achieved significantly higher choice prediction accuracy.