CLV Model Selection for Segmentation Perspective

Abstract

The customer lifetime value (CLV) metric aims to predict the importance level of each customer, offering to companies the ability to group them into homogenous segments, to propose appropriate marketing actions and to optimize resource allocation. CLV is widely suggested as a new base to segment customers. The Pareto/NBD and the BG/NBD are the most relevant CLV models, assuming that the number of transactions performed by customers follows a Poisson distribution. The BG/GCP has the particularity to model the number of transactions using the Conway–Maxwell–Poisson (CMP) distribution which is a generalization of the Poisson distribution providing additional flexibility when modeling discrete data. In this paper we propose to compare segmentation performance of the BG/GCP compared to the Pareto/NBD and the BG/NBD models, and to select the most efficient one. This performance is evaluated using three different clustering methods namely K-means, Fuzzy C-means and EM Clustering. Using two simulated datasets, presenting respectively an over and an under dispersion from Poisson distribution, the empirical analysis shows that the BG/GCP model based on CMP flexibility offers the best segmentation performance.