← Back to Glossary
Churn

Predict
Churn.

Updated

Reactive retention waits for the customer to click cancel and then tries to save them. Predictive retention identifies them weeks earlier — while there is still time and goodwill to intervene. For subscription businesses with meaningful customer counts, predicting churn is one of the highest-leverage analytical investments available.

The simplest way to predict churn

You do not need machine learning to start. A rule-based score using 4–6 signals catches the majority of at-risk customers:

  • Recent skipped cycle — +20 points.
  • Two skipped cycles in a row — +40 points.
  • Recent failed payment — +30 points.
  • Card expiring within 30 days — +25 points.
  • No portal login in 60+ days — +15 points.
  • Negative-sentiment support ticket — +30 points.

Sum across signals, anyone above a threshold (say, 50) gets flagged. Crude, fast, and surprisingly effective. Build it in a spreadsheet on a Monday and start using it Tuesday.

When to upgrade to machine learning

Move to ML-based prediction when:

  1. You have at least several thousand subscribers and a year of clean data.
  2. Your rule-based system is tuned and you want incremental lift, not a starting point.
  3. You have someone (in-house or consulting) who can maintain the model — not just build it once.
  4. You have measurable downstream interventions ready to consume the scores.

Don't skip the intervention design

The most common churn-prediction failure mode is a beautiful model with no follow-up. Before you build the prediction, design what happens to a high-risk customer: which save offer, which email, which manual outreach. The action is what generates the retention lift; the prediction just routes the customer to the action. See churn prediction model for the modeling side and win-back campaign for what happens after a customer does churn.

Frequently Asked Questions

How accurate is churn prediction?

A well-built gradient-boosting model typically reaches 75–85% AUC on subscription data. Rule-based scoring usually reaches 60–70% AUC. Anything dramatically higher is usually a sign of label leakage — features that are actually outcomes of churn rather than causes.

What signals are most predictive of churn?

For subscription commerce: recent skipped cycles, failed payment events, card expiration, negative-sentiment support tickets, and declining portal engagement. For SaaS: declining feature usage, drop in active users on the account, support escalations, and pricing-tier complaints.

How early can I predict churn?

For voluntary churn, reliable signals usually appear 2–6 weeks before cancellation. For involuntary churn, 30+ days of warning is normal (card expiration, prior failed payment). The further ahead the prediction, the lower the accuracy but the higher the intervention value.

What should I do with a churn prediction?

Trigger a relevant intervention — a save offer, a personal email, a phone call for high-value subscribers, a frequency-change suggestion, or a swap recommendation. The intervention design matters more than the prediction precision. A 70% accurate model paired with a great intervention beats a 90% accurate model with no follow-up.

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