| The goal of this dissertation is to compare two types of models for customer-base analysis in noncontractual settings, based on the predictions of future behavior and lifetime value. The first type of model, known as Lost-for-Good (LFG), assumes that all customer relationships have a finite endpoint and customers that drop out of a relationship will not return. The second type, known as Always-A-Share (AAS), assumes that customers move between different states of activity and never drop out for good. Much of the literature on noncontractual relationships is based on LFG models. These models are formulated to capture the unobserved attrition of customers. However, unlike AAS models the LFG models cannot incorporate time-varying covariates such as marketing variables or seasonality of demand. I propose a discrete hazard model, which is an AAS model, and compare its predictive performance against the Pareto/NBD model (Schmittlein et al., 1987), which is an LFG model. I use the hierarchical Bayes approach to estimate both the models.;The model comparisons are performed on transaction databases from three direct marketing companies that differ in the longitudinal and cross sectional characteristics of the demand. Two of the settings have high customer attrition, low transaction frequency, and seasonal demand. The other setting has low attrition and high transaction frequency. For the firms with high attrition the discrete hazard model performs better than the Pareto/NBD model based on measures of fit. For the firm with low attrition and high transaction frequency the Pareto/NBD performs better on the measures of fit. There is significant cross sectional variation in the fit of the models. In the case of the high attrition firms, the discrete hazard model does better than the Pareto/NBD for customers with low transaction frequency, while the Pareto/NBD does better for customers with high transaction frequency. I considered two alternative functional forms for the hazard, quadratic and log. The fit of these two functional forms is similar, with a slight edge for the quadratic specification. The empirical results indicate that the proposed discrete hazard model is a feasible alternative to the Pareto/NBD model.;Keywords: lifetime value, customer-base analysis, Pareto/NBD, Hazard models, Lost-for-Good models, Always-a-Share models... |
