Customer Lifetime Value and RFM Data: Accounting Your Customers: One by One

A customer behavior model that permits the estimation of customer lifetime value (CLV) from standard RFM data in "non-contractual" setting is developed by extending the hierarchical Bayes (HB) framework of the Pareto/NBD model (Abe 2008). The model relates customer characteristics to frequency, dropout and spending behavior, which, in turn, is linked to CLV to provide useful insight into acquisition. The proposed model (1) relaxes the assumption of independently distributed parameters for frequency, dropout and spending processes across customers, (2) accommodates the inclusion of covariates through hierarchical modeling, (3) allows easy estimation of latent variables at the individual level, which could be useful for CRM, and (4) provides the correct measure of errors. Using FSP data from a department store and a CD chain, the HB model is shown to perform well on calibration and holdout samples both at the aggregate and disaggregate levels in comparison with the benchmark Pareto/NBD-based model. Several substantive issues are uncovered. First, both of our datasets exhibit correlation between frequency and spending parameters, violating the assumption of the existing Pareto/NBD-based CLV models. Direction of the correlation is found to be data dependent. Second, useful insight into acquisition is gained by decomposing the effect of change in covariates on CLV into three components: frequency, dropout and spending. The three components can exert influences in opposite directions, thereby canceling each other to produce less effect as the total on CLV. Third, not accounting for uncertainty in parameter estimate can cause large bias in measures, such as CLV and elasticity. Its ignorance can potentially have a serious consequence on managerial decision making.