A Bayesian implementation for predicting and understanding customers’ purchase behavior

Abstract

In this project, we implement several Bayesian models of Customer Base Analysis and test them on a real dataset of customer transactions from a wholesales company. Our goal is to predict the amount of purchases that customers will make in the future, as well as understanding the hidden dynamics that drive their purchase patterns. After describing the particularities of the dataset, we introduce the traditional Pareto-NBD model by Schmittlein et al.(1987) and the extensions proposed by Abe (2009), which make it possible to expand the assumptions of the original model and to incorporate new features. We estimate the parameters of these models with Full Bayes and compare their performance. Next, we introduce the recent Gaussian Process Propensity Model (Dew and Ansari, 2018) and propose an adaptation in order to model count data. This model is also implemented in a Bayesian fashion and validated on the given dataset.