Software Growth

Churn prediction

Churn prediction uses past behavior to flag accounts likely to cancel, so you can act first. It ranges from simple usage alerts to machine learning models.

Churn prediction means finding the customers who are likely to leave while they are still paying. Churn is a lagging number: by the time it shows up in your churn rate, the customer is gone. Prediction looks at behavior that came before cancellations in the past and watches for the same patterns now.

What to look for

The strongest signals in most SaaS products are about use. Sequoia's data science team puts it simply in its retention essay: engagement drives retention. So a drop in engagement comes first. Typical warning signs:

  • Fewer logins or active users than the account's own recent average.
  • Core features no longer used, or key setup steps never finished.
  • The main contact leaving the company or going silent.
  • Support tickets that stay unresolved, or a run of complaints.
  • Failed payments or a payment method about to expire.
  • Seats removed, or a downgrade request.

Three levels of sophistication

  1. Rules. "No login for 21 days" or "usage down by half versus last month". Easy, transparent and often enough.
  2. Scoring. Combine several signals with weights into a customer health score.
  3. Models. Train a statistical or machine learning model on past churn to output a probability per account. This needs hundreds of churn events and clean data. Gainsight, for one, argues for data-driven scoring over gut feel in its health score post.

How to check that it works

Recall measures the other side: how much of the real churn you caught.

Say you flag 25 accounts this quarter and 10 of them cancel. Precision is 10 / 25 = 40%. If 20 accounts churned in total, recall is 10 / 20 = 50%. You caught half the churn, and 4 in 10 of your alerts were right. Test any rule on last year's data before you rely on it.

Prediction is not retention

A flag has no value until someone acts. Decide what happens for each type: a personal email, a call, a training offer, a fix for a missing feature. Then measure whether flagged accounts that got the action stayed more often than those that did not. Customers flagged because of a failed payment need dunning, not a conversation.

For small SaaS

With under 500 customers, a model will not beat a sorted list. Export last login date per account every Monday, read the bottom ten, and write to them. Keep notes on what the ones who churned looked like in their last 60 days, and that becomes your rule set.

Sources

Back to the SaaS glossary