Wrap-up#

Congratulations if you made it until the end :) This tutorial was rather long and technical, but gives you an idea about how complex this could get to create a churn model. We saw some of the major difficulties:

  • defining properly what is churn and how to translate it into a “target” variable

  • creating features depending on time, and relatively to a reference date

  • choosing the proper cross-validation strategy

Of course, handling this kind of models in real life will take you much further. It is very frequent to include hundreds of features, or even spend a few days trying to properly scope the churn definition.

And this might only be a first step. Indeed, you’ll probably want to model the propensity to respond to different kind of anti-churn offers, or model the true incremental effect of these systems (“lift modeling”).