The Smallest Possible Health Model
Three inputs beat twelve: activation state, usage recency/depth, and key feature adoption. Start with equal weights. If AUC/ROC against churn is ~0.5, your inputs are wrong, not your math. Fix the signals first.
-
-
Health Scores: Fewer Signals, More Signal
Health scores work when they are legible and predictive. Keep the inputs few, stable, and behavior-based (e.g., activation milestones met, usage depth/recency, key feature adoption). Write the attribution rules down so you can explain changes. If you can’t predict churn or expansion better than chance, your model is a vanity metric—fix the inputs before tuning…
-
Experience and Outcomes are all that matter.
Churn is a lagging indicator. By the time a customer cancels, the problem started months earlier, usually as low engagement or stalled adoption. If you’re reacting to churn risk at renewal time, you’re already behind. The simplest formula for keeping customers I have been able to validate is below: Retention = Experience + Outcomes(Adoption(Engagement)) Work…