RFM 分析で使われるリーセンシー(直近の購買からの経過時間)とフリークエンシー(購買頻 度)のデータから、一般的な消費者行動の仮定に基づいて、ある時点での顧客の生存確 率を推定する。既存の経験ベイズに基づいたPareto/NBD モデルを階層ベイズの枠組 みに改良し、購買率と離脱率を表すパラメータに共変量を組み込むことによって、マ ーケティングに有益な知見が得られる。実証研究として、日米２種類の顧客購買デー タを使い、このモデルを評価する。

This research extends a Pareto/NBD model of customer-base analysis using a hierarchical Bayesian (HB) framework to suit today's customized marketing. The proposed HB model presumes three tried and tested assumptions of Pareto/NBD models: (1) a Poisson purchase process, (2) a memoryless dropout process (i.e., constant hazard rate), and (3) heterogeneity across customers, while relaxing the independence assumption of the purchase and dropout rates and incorporating customer characteristics as covariates. The model also provides useful output for CRM, such as a customer-specific lifetime and survival rate, as by-products of the MCMC estimation. Using two different types of databases --- music CD for e-commerce and FSP data for a department store, the HB model is compared against the benchmark Pareto/NBD model. The study demonstrates that recency-frequency data, in conjunction with customer behavior and characteristics, can provide important insights into direct marketing issues, such as the demographic profile of best customers and whether long-life customers spend more.