← Back to Glossary
Churn

Churn Prediction
Model.

Updated

Churn prediction turns reactive cancel-flow optimization into proactive retention. Instead of waiting for a customer to click cancel, you flag them weeks earlier — while there is still time and budget to intervene. Done well, it is one of the highest-ROI uses of customer data a subscription business can build.

What signals go into a churn model

  • Engagement decay — declining email opens, fewer logins, longer gaps between active sessions.
  • Subscription actions — pause requests, frequency changes, skipped cycles, swap requests.
  • Payment events — failed charges, card updates, dunning emails opened.
  • Support friction — tickets opened, sentiment in messages, time-to-resolution complaints.
  • Tenure and cohort — month-since-signup, signup channel, signup discount applied.

Common modeling approaches

  1. Rule-based scoring. The starting point: assign points for risky behaviors (skip + skip + support ticket = high risk). No ML required, surprisingly effective for early-stage operators.
  2. Logistic regression. Statistical model that outputs a churn probability per customer. Easy to interpret — you can see which features drive the score.
  3. Gradient boosting (XGBoost, LightGBM). The workhorse of production churn prediction. Higher accuracy than logistic regression, harder to explain to non-technical stakeholders.
  4. Survival analysis. Models time-to-churn rather than binary outcome. Useful for forecasting cohort revenue, not just flagging at-risk individuals.

What to actually do with predictions

A score is useless without an intervention. The best churn-prediction setups feed directly into automated retention plays: high-risk customers get a personal email from a success manager, a discount offer, a swap suggestion, or a phone call. For most subscription merchants, a simple rule-based model plus a tailored cancel flow recovers more revenue than a sophisticated ML model with no follow-up. Start with the action, then build the model that triggers it. See customer churn modeling for the broader analytical view.

Frequently Asked Questions

Do I need machine learning to predict churn?

No, especially at small scale. A rule-based scoring system using 4–6 signals (skip behavior, support tickets, engagement decay, payment failures) catches most at-risk customers and is faster to build than an ML model. Move to ML once you have thousands of customers and clean data.

What is the accuracy of a typical churn prediction model?

A well-built gradient-boosting model typically achieves 75–85% AUC on subscription data — meaning it correctly ranks at-risk customers above safe ones in that share of pairs. Higher accuracy is often a sign of label leakage rather than a better model.

How far in advance can churn be predicted?

Reliable signals usually appear 2–6 weeks before cancellation for voluntary churn. Involuntary churn is more predictable — card expiration and prior failed payments give 30+ days of warning. The further ahead you predict, the lower the accuracy but the higher the intervention value.

What data do I need to train a churn model?

At minimum: subscription event history (signups, cancels, pauses), billing events (successful and failed charges), engagement signals (email opens, portal logins), and at least 6–12 months of data so the model sees full churn cycles. Cleaner data beats more data.

Start Growing Your Subscription Revenue

Join 5,000+ Shopify merchants using Joy Subscriptions. Free to install, no credit card required.

  • Free 14-Day Trial
  • No Credit Card Required
  • Cancel Anytime