What Thai prepaid patterns teach postpaid churn models
Postpaid teams often borrow prepaid drought charts because the pictures look related. They are not. This note is the argument we make in Telco Churn Signals (Thailand).
Prepaid silence in Thailand is a cash-flow event. A subscriber can vanish for three weeks, revive on payday, and still be loyal in any sense that finance would recognise. Postpaid silence is usually a mandate still clearing while attention has already left. If you train one model on both, the payday revival teaches the postpaid side to wait, and the postpaid inertia teaches the prepaid side to panic.
Drought is a state, not a decay
A recency decay assumes the world is smooth. Top-up behaviour is not. In clinic we encode drought as consecutive dry weeks against the customer’s own median gap, not against a national average. A tourist SIM with one top-up and a quiet month is not the same object as a Bangkok commuter who tops up every Friday and then misses two.
When postpaid teams copy decay anyway, they invent “risk” in every holiday week. Songkran is not a mass resignation from mobile service. It is a calendar.
SIM rotation and the subscriber key
People rotate SIMs for coverage, for promotions, and for reasons that will never appear in your CRM. If your warehouse cannot stitch a person across MSISDNs, your churn rate is partly a counting error. We do not magically solve identity in a four-week course. We do force the slide to say “line expiry” when that is all you can observe.
Tourist lines and apparent mass exit
Arrivals into Suvarnabhumi and Don Mueang create prepaid cohorts with short natural lives. If those lines sit in the same extract as resident postpaid, your monthly churn spike may be an arrivals chart. Segregate first. Model second.
What postpaid should actually steal
Not the drought toy. Steal the discipline of not collapsing labels. Prepaid operators already distinguish expiry, port-out, and involuntary non-reload. Postpaid slides that dump voluntary cancel and failed mandate into one “churn %” are less mature than a street kiosk report.
Also steal the humility about family plans. A quiet secondary line is not a save-desk emergency. Prepaid multi-SIM households taught us that years ago; postpaid CRM still lights up as if every quiet MSISDN were a departing breadwinner.
A small experiment we run in lab
Take a desensitised mix, train a single binary model, then train two. Compare not AUC but disagreement with the save queue in a campaign week. The split model is usually worse on a leaderboard and better on the floor. That is the point of Churn Prediction Analytics as we teach it: the floor wins.