Churn Prediction Analytics as a four-layer desk, not a single score.
This page is the studio’s working map. Courses drill into each layer; the map keeps finance, data science, and the save queue from inventing three different definitions of “gone.”
A
Label layer
Exit is split: voluntary cancel, involuntary non-pay, and dormancy. Family plans and multi-SIM households are not flattened into one subscriber id just because the warehouse prefers it. If finance cannot sign the label, the model does not start.
B
Window layer
Features follow invoices or top-up cycles, not Monday-to-Sunday convenience. Prepaid drought is a state with duration, not a decaying recency toy. Postpaid lookbacks stop at the bill boundary unless you can explain why they should not.
C
Decision layer
A score without capacity is trivia. Thresholds are set against the number of conversations a Bangkok save desk can actually hold this week, including campaign weeks when the queue is already full.
D
Watch layer
After launch we track silent feature breakage, calendar shocks, and disagreement between desk and model. The artefact is a one-page sheet, not a twelve-tab dashboard nobody opens.
Where Thai operations press on the grid
Prepaid top-up droughts can look like death and then revive after a payday. Tourist SIMs inflate apparent churn if your extract cannot separate them. Festival weeks move recharge volume without moving intent. The Grid method treats those as first-class constraints, which is why Telco Churn Signals (Thailand) exists as a separate syllabus rather than a footnote in a generic MOOC.
What we refuse to call analytics
A leaderboard metric with a leaked label. A “next best action” slide that ignores agent handle time. A black-box vendor score with no window diagram. Systemagentgrid will teach you to recognise those and walk away.