Predicting the inactivity and the repeat transaction frequency of a firm's customer base is critical for customer relationship management. The literature offers two main approaches to such predictions: stochastic modeling efforts represented by Pareto/NBD and machine learning represented by neural network analysis. As these two approaches have been developed and applied in parallel, this study systematically compares the two approaches in their prediction accuracy and defines the relatively appropriate implementation scenarios of each model.
By designing a rolling exploration scheme with moving calibration/holdout combinations of customer data, this research explores the two approaches' relative performance by first utilizing three real world datasets and then a wide range of simulated datasets.
The empirical result indicates that neither approach is dominant and identifies patterns of relative applicability between the two. Such patterns are consistent across the empirical and the simulated datasets.
This study contributes to the literature by bridging two previously parallel analytical approaches applicable to customer base predictions. No prior research has rendered a comprehensive comparison on the two approaches' relative performance in customer base predictions as this study has done. The patterns identified in the two approaches' relative prediction performance provide practitioners with a clear-cut menu upon selecting approaches for customer base predictions. The findings further urge marketing scientists to reevaluate prior modeling efforts during the past half century by assessing what can be replaced by black boxes such as NNA and what cannot.
