12 March 2026
Hazard windows are not “the next 30 days”
In the Early Warning Cohort we still meet people who inherited a tile labelled “churn risk — 30 day”. The number is round, the SQL is copy-pasted, and finance understands months. None of that makes it a detection window.
A hazard window is the stretch of time in which the chance of a cancel (or a true lapse) is being estimated, given that the person has survived until the start of that stretch. It has a beginning that you can defend, an end that you can defend, and a rule for who is not in the risk set. A rolling calendar month has a beginning that is “whenever the dashboard refreshed”. That is a filing habit.
The gym week that never counted
A membership club on the Gower showed us eight accounts that always went quiet in the third week of the month, when overtime peaked at the nearby plant. The 30-day flag lit up every time, then went dark after payday visits. The save team spent Thursdays apologising for emails that should never have sent. The hazard we marked instead began after the person’s own quiet week had passed without the usual return, not after an arbitrary Monday.
What we write on the brief
Start date: first day after expected cadence is missed, plus a grace that matches the product (two sessions for a climbing wall, one missed collection for a veg box). End date: the last day a save action is still cheaper than replacing the member. Exclusions: pauses with a recorded reason, cooling-off cancellations, and deaths of the contract that are admin, not sentiment.
If you cannot write those three sentences, you do not have Churn Detection Analytics. You have a countdown.