Half the time to churn

Measure the average time between an account first showing decline and the account leaving. That is your window. Your detection has to operate at roughly half of it, and most operations detect at somewhere near the end.

The reasoning is not subtle. Detecting risk at ninety percent of the way to churn leaves you the remaining ten percent to diagnose, act, and have the action take effect, and the customer has spent the other ninety percent becoming more certain.

Getting the number requires looking backward at accounts that already left, which is uncomfortable and is the only place the data exists. Take twenty churned accounts and find, in usage, support and engagement data, the point at which the pattern first changed. Not the point at which somebody noticed. The point at which it changed.

You will usually find that the change was visible in data you already collected, six months out, and that nothing was watching it.

Then set the detection threshold at half the window and accept that it will produce false positives. A program tuned to avoid false positives will detect at the end, because certainty arrives late. It is cheaper to have a conversation with a healthy account than to miss a leaving one.