A New Method for Evaluating Stochastic Models of Brand Choice

Abstract

This article explores the use of a brand choice stochastic model's mean value function in evaluating two models empirically, using a common set of purchase data. The linear learning model fit the data well, but its mean value function was not capable of making reasonable predictions of successive, aggregate purchasing statistics. Another brand choice model, the new trier model, was found to perform much better. The results suggest that model tests should not be restricted to the usual goodness-of-fit test, especially in situations of non-stationarity. A structural comparison of the two models focuses on their different approaches to nonstationarity.