Customer Purchase Frequency Modeling Using Transactional Data and Regression
Abstract
This review examines regression models for customer purchase frequency using transaction data. It explains how transaction records are aggregated into count outcomes and behavioral predictors, including recency, frequency, monetary value, promotion exposure, and inter-purchase intervals. The review compares ordinary least squares, Poisson regression, negative binomial regression, zero-inflated models, hurdle models, and the Conway-Maxwell-Poisson distribution. It shows that overdispersion, excess zeros, customer heterogeneity, observation-window design, and unequal exposure periods strongly affect model choice and interpretation. Representative studies indicate that negative binomial and two-part count models often fit purchase data better than simple Poisson or linear models, although results depend on data construction and validation procedures. The review also summarizes common evaluation measures, including mean absolute error, root mean squared error, deviance, and information criteria. Current limitations include inconsistent definitions, censoring, temporal leakage, non-stationarity, privacy concerns, and limited causal interpretation. Future research should develop dynamic and seasonal count models, Bayesian uncertainty estimation, causal analysis, and interpretable machine-learning approaches.