Notes from the retention desk
Short pieces on what makes churn prediction analytics trustworthy for app teams — and what quietly breaks them.
Leading Indicators That Appear Before Cancellation
Most teams watch churn after the fact. Here is how UK product squads spot drift two to four weeks earlier — without inventing a dozen new metrics.
Labelling Churn Without Poisoning Models
Voluntary cancel, involuntary lapse, and long dormancy are not the same outcome. Treat them as one label and your predictions will lie politely.
When Not to Build a Churn Model Yet
Sparse events, unstable plans, and unowned interventions make fancy scores useless. A readiness check saves months.
Calibrating Save Offers to Risk Bands
Blanket discounts train bargain hunters. Tiered responses matched to churn-risk bands protect margin while lifting retention.