Silent accounts: when usage drops before the cancellation form
A mandate can keep clearing while the product has already been abandoned. If you call that voluntary churn, your save-desk will look heroic and your risk model will hunt the wrong people.
In the teaching extract we keep a stubborn population: near-zero usage, successful charges, no cancel ticket. Marketing wants them in the churn denominator when they finally stop paying. Finance wants them in revenue until the mandate dies. CRM wants a campaign. Data science is asked to “just predict churn” as if those three sentences were one fact.
Dormancy is a label, not a feature
If you treat silence as a predictor of cancel, you will be partly right — many silent accounts do eventually stop paying. You will also spam people who intended to keep a backup subscription, a parental control line, or a tool they use once a quarter. The Grid method puts dormancy in layer A (label), not only in layer B (features).
Operationally that means three scores or at least three thresholds: likely voluntary exit, likely involuntary non-pay, likely dormant-but-paying. One probability dumped into one queue is how agents learn to ignore you.
How silence shows up differently in Thailand
Prepaid silence is often cash timing. Postpaid silence on a family plan is often a secondary line. Insurance affinity programmes — we see these in Forecast Floor work — accumulate “just in case” policies that look dormant and are doing exactly what the customer bought. Copying a SaaS “days since last login” heuristic across those worlds is not analytics. It is costume.
What we ask students to produce
A one-page definition: how many days of near-zero usage, on which products, excluding which pause SKUs, before an account is called dormant. Finance initials it. Only then may recency of usage enter a voluntary-exit model, and only with a window that stops before the label timestamp.
Students who skip the initial and jump to a classifier usually “win” a metric and fail the queue rehearsal in week eight. That failure is inexpensive in lab and expensive in production. We prefer it happens on Chan Rd. time, in a fourteen-seat clinic, while the extract is still desensitised.