Unfortunately, the Pareto-NBD log-likelihood function cannot be estimated as implemented in that package if you have customers with rich purchase histories – in the hundreds of items – which you do if you’re a popular e-commerce platform that has been around for a while and it has saved its RFM data diligently.
The reason for this is explained here and there is a fix too, based on a computational trick explained here. I am not sure why the author never suggested it as a formal patch to the original package. Maybe it’s in the works. Until then, my patched version of the library is here.
There’s still a limitation. Though this fix allows you to estimate the four parameters of interest from a sample of customers that includes people with rich purchase histories, individual probabilities that such customers are still alive cannot be estimated: the
pnbd.PAlive() function will return
This may not be a deal-breaker for you. There’s room for a judgment call here, based on the increasing frequency paradox described in the walkthrough. Any marketing campaign should target newer or less engaged customers. For these,
pnbd.PAlive() still works. Customers with rich histories who haven’t been seen in a long time are probably dead, so there’s no use worrying about them, while very active customers who have been seen very recently are best not bothered.