Application of Pareto/NBD and Pareto/GGG models for customer lifetime value estimates in high churn setting I have been attempting to estimate customer lifetime value in the context of online classifieds (high churn context) using probabilistic models, chiefly the Pareto/NBD and Pareto/GGG techniques available through the 'BTYD' and 'BTYDplus' packages in R.
I constructed a cohort of users and tracked their behaviour over time and despite what appeared to be impressive results in the holdout period (~2% deviation), the diagnostics plots (incremental and cumulative) showed marked deviation compared to the results shown in many of the tutorials.
It would be great if someone is able to offer any advice or suggestions to improve these estimations in a high churn setting. Are there alternative techniques that are better suited to such a business setting?
Model performance in holdout period
Diagnostic plots of Pareto/GGG model
Thanks in advance for any help you are able to offer.
 A: Maybe if you explain the context of online classified it would help to provide guidance.
According to the paper by Fader@Hardie 
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.180.2024&rep=rep1&type=pdf
It seems that a crucial part of the application of Pareto/NBD is the previous assessment of the axis of non-contractual-contractual/discrete-continuous type of business. Depending on what you find is the proper course of action (maybe Pareto/NBD is just not the most suitable one)
A: I don't know if that's the case for the BTYD library, but there can be a penalizer coefficient that helps control the parameters of the model's fit, not letting them get too big.
That's implemented in the lifetimes Python library. Maybe that can help.
A: You might want to have a look at the implementations provided by the package CLVTools (https://github.com/bachmannpatrick/CLVTools). Many have been designed to work with datasets that are characterized by a high number of zero-repeaters, i.e. customers that only buy once. I have been succesful in applying these models for zero-repeater percentages as high as 80-90%.
You might find the vignette of this package helpful:
https://cran.r-project.org/web/packages/CLVTools/vignettes/CLVTools.pdf
