Twelve early-warning signals that actually precede subscription exit

Most “churn feature” blog posts start with recency of login and stop when the notebook looks busy. In clinic we ask a ruder question: would this field have been knowable before the customer had already decided to leave, and would an account manager in Yan Nawa know what to do with it in four minutes?

Below are twelve signals we keep seeing ahead of voluntary cancel or prepaid expiry. Each one is paired with a failure mode. If you copy the list into a model without the failure modes, you have not followed the lab.

Usage shape, not just recency

1. Session compression. The same weekly minutes, squeezed into fewer days. Often precedes cancel on streaming. Fails on shift workers whose pay cycle already compresses viewing.
2. Feature abandonment inside the product. A paid add-on that goes untouched for two bill cycles. Useful. Dangerous if the add-on was auto-bundled and never wanted.
3. Device drop-off. A second household device goes quiet. On family plans this is often a teenager changing phones, not an exit.
4. Support intensity without resolution. Ticket volume rising while CSAT stays flat. The leaked version of this field includes tickets opened after the save desk called. Audit the timestamp.

Money and contract texture

5. Discount cliff. A promotional rate ending inside the next invoice. Strong on postpaid. Weak if your warehouse backfills the “would-be” price after the bill is issued.
6. Failed tender, then success. One failed charge followed by a successful retry. Involuntary non-pay is a different label; do not fold it into voluntary risk.
7. Plan downgrade that does not reduce usage. People sometimes shrink the plan before they leave. People also shrink the plan because they finally read the invoice. Context first.
8. Top-up drought as a state. For prepaid, count consecutive dry weeks, not a smooth recency decay. Payday revivals will punish a decay toy.

Relationship and calendar

9. Save-desk contact in the prior quarter. Prior saves predict future saves and, sometimes, exhaustion. Never include the outcome of the current save attempt.
10. Campaign fatigue. Three or more outbound attempts in a billing window with no inbound reply. Capacity matters: your model should not flood the same queue.
11. Festival-week inversion. Usage up, recharge up, intent unchanged. If your score spikes every Songkran, you modelled the calendar.
12. Silent account before the form. Usage near zero while the mandate still pays. This is dormancy. Treat it as its own label; see the November journal piece.

How we use the list in lab

Students pick four signals, draw windows against the bill cycle, and must retire one of them after the leakage audit. The retired signal is usually number 4 or 5. That retirement is the learning, not the remaining AUC.

If you want the worksheets that accompany this list, they ship with Signal Pass and the flagship lab. The journal is free; the critique is not.

Back to the journal · Cohort Retention Lab