Labelling Churn Without Poisoning Models

A single "churned" flag mixes people who walked away, people whose card failed, and people who never returned after a quiet month. Models trained on that mix learn the average of three different stories — and recommend the wrong intervention for each.

Write explicit definitions: voluntary cancel within N days, involuntary lapse after failed renewals, and dormancy after M days without a meaningful session. Keep them as separate labels or nested flags. Document the rules next to the event dictionary so future releases do not redefine churn by accident.

Teams that treat labels as a living contract — with owners and review cadence — keep prediction work honest long after the first model ships.