Hello, I am following the CLV Quickstart tutorial and fitting BG/NBD Model and Gamma-Gamma model on transaction data for a project. It’s working great, thank you for this amazing library!
I am wondering: is it possible to extend the BG/NBD and Gamma-Gamma models to incorporate additional regressors? For example, if I have additional customer information (i.e. customer location, gender, attribution channel, etc.) that I believe will improve predictions on customer dropout or LTV, can I incorporate these additional regressors in addition to the usual frequency/recency regressors? Can you point to any resources on how to accomplish this?
The Pareto/NBD model currently supports time-invariant covariates, which would cover many of the variables you mentioned. It can be used for all the same use cases as BG/NBD.
There is a discussion in the pymc-marketing repo with additional information:
@ColtAllen you mentioned that there are plans for other models to support additional covariants in the future (assume you mean the BG/NBD and Gamma-Gamma Models)? Is there a pull request outstanding or anywhere else I can track progress on this? Thanks again.
There are open repo issues with research links; just filter by the CLV label.
The Gamma-Gamma model does not support covariates, but the BG/NBD and Pareto/NBD models can support both time-invariant covariates and high/low seasonality, and the Shifted Beta-Geo model can be modified to support time-varying covariates. I’ve also found two research papers outlining how to build a Beta-Geo model with time-varying covariates, but combining both approaches into a holistic model will be an applied research effort in itself.