Predicting Churn 45 Days Before It Happens: A Practical Guide
Most churn prediction models fail because they're trained on the wrong signals. Here's the framework we use to build accurate early-warning systems — and the data you actually need.
Churn is rarely a surprise. In hindsight, the signals were almost always there. The challenge is detecting them early enough, across hundreds or thousands of accounts, to actually intervene.
Here's the framework we use.
The Signals That Actually Predict Churn
Most churn models are built on obvious lagging indicators — login frequency, support ticket volume. These are too late. The signals that matter are subtler:
Engagement decay: Not just "did they log in" but "are they using the features that correlate with long-term retention?" Feature-specific usage often predicts churn 60+ days out.
Sentiment drift: Support tickets, NPS scores, and customer communications contain tone signals. Customers who are planning to churn often shift to a more transactional tone weeks before they submit a cancellation.
Champion changes: When your internal champion at a customer leaves the company, churn probability spikes. Monitoring LinkedIn or CRM contact activity can surface this.
Billing friction: Failed payments, downgrades, and discount requests are obvious. But also look at delayed invoice approvals — procurement slowdowns often precede cancellation.
Building the Model
You need at minimum 12 months of historical data with known outcomes. For each churned customer, label the 90-day window before churn and extract the signal features. Train a gradient boosted model (XGBoost works well for this).
The model should output a churn probability score updated daily. Flag any account that crosses a threshold for CSM review.
The Intervention Playbook
An early warning is only valuable if it triggers action. The AI CS agent should:
- Detect the risk flag
- Draft a personalised intervention brief for the CSM: what signals fired, account context, suggested talking points
- Trigger an automated touchpoint if the CSM can't reach the account within 48 hours
With this system, we typically see churn reduction of 30–50% within 90 days of deployment.