Predicting who will leave.
WHAT MUST BE DEFINED FIRST
What churn means.
WHAT THE OPTIONS ARE
Explicit cancellation No activity for a defined period Subscription lapsing without renewal Spend falling below a threshold
WHY THE DEFINITION CHANGES EVERYTHING
Each produces a different target, different data, and a different model.
WHAT THE PREDICTION WINDOW MUST BE
Long enough to act, short enough to be accurate.
WHY THAT BALANCE MATTERS
Predicting churn the day before it happens is useless.
WHAT FEATURES TYPICALLY MATTER
Recency and frequency of use Trend in usage, not only level Support contacts and their sentiment Payment failures Tenure Product breadth
WHY TREND MATTERS MORE THAN LEVEL
Declining usage predicts departure; low but steady usage frequently does not.
WHAT LEAKAGE TO WATCH FOR
Cancellation-related activity recorded before the formal cancellation date.
WHAT TO DO WITH PREDICTIONS
Target intervention at those where intervention plausibly helps.
WHY THAT QUALIFICATION
Some customers leave for reasons no offer changes, and discounting them is pure cost.
WHAT TO MEASURE
Retention among those contacted against a held-back group.
