Churn intervention loop

"We have a churn score. Nobody does anything with it."

The existing model predicted churn accurately and was useless: it ranked customers by risk, not by how much an intervention would change their behaviour. Retention spend went to people who were leaving regardless.

An uplift model over the same features, a treatment assignment service with a permanent holdout, and an orchestration loop that picks the intervention, respects contact fatigue rules, and measures the treated-versus-holdout gap weekly.

Marketing wanted to override the holdout for a quarterly push. We kept the holdout and made its cost explicit on the dashboard instead of arguing about it.

Telco Customer Churn (IBM) + Hillstrom e-mail uplift trial

Two notebooks: why a risk model misdirects spend, and the same data modelled for uplift. The assignment service is in the repo with the holdout logic intact.