How to build an early-warning system for member churn

How to build an early-warning system for member churn

By the time someone leaves, the decision was made weeks earlier. Leaves are a lagging indicator. Here's how to build a leading one from behavior that precedes the exit.

Step 1: define pre-churn signals
From members who left, look backward: most show declining posting frequency and lengthening gaps for 2-4 weeks before leaving. The signal is the trend, not any single quiet day.

Step 2: build a simple risk flag
Flag members whose weekly activity dropped to a fraction of their personal baseline for two consecutive weeks. Personal baseline matters — a once-a-week poster going silent is different from a daily poster going silent.

Step 3: act proportionally
Don't ping flagged members with 'you ok?' — that's creepy and accelerates exits. Instead, route them content they once engaged with, or surface them in channels they used to frequent.

What the evidence supports
Declining-frequency-before-leaving is a robust pattern across subscription products. The hard part is precision — most people who slow down don't leave, so any flag has high false positives.

The caveat
A churn predictor with 80% false positives is worse than useless if you act intrusively on it. Use it to inform soft, ambient nudges only.

Open question: if intervention risks accelerating exits, is the right move to predict churn at all — or to just build a community good enough that the baseline holds?