Improving the Forecasting Accuracy of 2-Step Segmentation Models

The estimation of consumer preferences with choice-based conjoint (CBC) models is well-established. In this context, the use of Hierarchical Bayesian (HB) models, which estimate consumers’ individual preferences is nowadays state-of-the-art. However, the knowledge of consumer preferences on a less disaggregated level, like segment-level, is key for demand predictions of non-customized products. Clustering individual HB data to achieve segment-level preferences is known as inappropriate, since 2-step segmentation approaches generally underlie 1-step approaches, e.g., Latent Class models. But, may the inclusion of different concomitant variables into the clustering process of individual CBC data relax that disadvantage? To answer this question, we used an empirical data set and compared the forecasting accuracy of 1- and 2-step approaches. While demographic variables showed small effects, psychographic variables turned out to heavily improve forecasting accuracy. In particular, 2-step approaches, that consider psychographic variables within the clustering process, showed a forecasting accuracy comparable to the one of 1-step approaches.