We adopt continuous mixture models, specifically,
the beta-geometric (BG)/NBD model (Fader et al.,
2005a). We prefer this approach to the Pareto/NBD
model (Schmittlein et al., 1987) because it creates fewer
estimation problems and offers clear interpretations.
The BG/NBD approach considers two processes,
purchase frequency and the ‘‘death’’ of the customer,
and it makes two fundamental assumptions about
purchase frequency and inactivity. For frequency (i.e.,
distribution of the number of purchases and heterogeneity
across customers), the number of purchases Y
follows a Poisson distribution with a transaction rate k.
The probability that a customer makes y transactions
during time t is PðY ¼ yÞ ¼ ðktÞy
y! ekt: This expression
means the interpurchase time follows an exponential
distribution. The process is withoutmemory, and there
is a constant probability of a transaction in each period.
The market is fundamentally stationary. These
assumptions have been well validated for frequently
purchased goods (Ehrenberg, 1959, 1988).