Before the notion of "customer lifetime value" became a popular phrase among managers, database marketers were using simple notions to assess the value of different customer groups in relation to their past behavioral patterns. The most popular framework classifies prospects based on RFM: the recency, frequency, and monetary value of past transactions.
We present a stochastic model that formalizes some of these ideas in order to link the recency and frequency of past transactions with valid estimates of future activity. Specifically, we allow for a Poisson changepoint process with two (nonhomogeneous) regimes. We allow for heterogeneity in all three components of the model ("early" behavior, "steady state" behavior, and the probability of switching at any point in time).
We show that a limited amount of initial transaction data is sufficient to uncover accurate forecasts of steady state purchasing and, thus, we can obtain valid estimates of future lifetime values based on observed summary statistics such as recency and frequency.
|