Churn is when a customer stops doing business with you: a subscriber cancels, a gym member doesn't renew, a borrower moves to another lender. Churn prediction uses your own data to estimate, for each customer, how likely they are to leave in the coming weeks — early enough to do something about it.
This guide explains what a churn model needs, why scores without reasons are of little use, how to act on them, and how to check the model honestly. To see risk scores with their reasons, try the customer churn prediction demo.
Define "leaving" first
It sounds obvious, but it decides everything:
- Subscriptions and broadband: a cancellation, or not renewing.
- Gyms: no renewal within 30 days of expiry — or no visit in 45 days, which comes first.
- Retail and cafés: no purchase in a set period, which depends on how often regulars normally buy.
- Lenders: closing a loan early and not taking another, or moving a credit line elsewhere.
Pick a definition the business already uses, and a time window — "likely to leave in the next 60 days" — that leaves time to act.
The data a churn model uses
Most of it you already have:
- Usage: logins, visits, data used, orders — and especially how they are changing
- Service: complaints, tickets, outages, delivery problems
- Payments: late payments, failed auto-debits, downgrades
- Tenure and plan: how long they've stayed, what they pay, recent changes
- Contact: responses to messages, offers used
A useful rule of thumb: you need at least a year of history and a few hundred customers who actually left, so the model can learn what leaving looks like.
Every score needs reasons
A list that says "Meera: 81% risk" helps a little. A list that says "Meera: 81% — usage down 60% this month, likely moving house" tells the team what to do. Modern models can show which factors pushed each customer's score up, in plain words.
Reasons also catch bad models: if the top reason is something meaningless, the model has learned the wrong thing.
Turn scores into actions
Different reasons call for different actions:
| Main reason | Better action than a discount |
|---|---|
| Slow service or outages | A technician visit or a fix, with a call |
| Usage falling | A check-in call or a plan that fits better |
| Price comparison | A loyalty offer or an upgrade with more value |
| Moving away | A transfer to another branch or city |
| Late payments | A flexible payment date |
Blanket discounts at renewal cost money on customers who would have stayed anyway. Acting on the reason costs less and works better.
Check that it works — honestly
- Test on unseen customers. Measure the model on customers it didn't learn from.
- Two numbers matter: of the customers who left, how many did the model flag in advance? And of those it flagged, how many really were at risk?
- Re-check every month. Compare last month's scores with who actually left, and retrain when behaviour changes.
- Measure the actions too. Keep a small group of flagged customers with no action, to see whether the calls and offers make a difference.
Be wary of anyone who quotes an accuracy figure before seeing your data. It depends on your customers, your data and your definition of leaving.
Privacy
Churn models use personal and behavioural data. Keep them on your own servers or a private cloud, use only the data the purpose needs, limit who sees individual scores, and follow the Digital Personal Data Protection Act's rules on notice and purpose.
Where to start
- Agree the definition of leaving and the prediction window.
- Pull 12–24 months of history for usage, service, payments and plans.
- Build a first model and look at its reasons with the team that talks to customers.
- Score every week; act on the high-risk list; keep a hold-out group.
- Review monthly, and improve the data and actions before the model.
The short version
Churn prediction gives you a weekly list of customers likely to leave, with the reasons, while there's still time to keep them. It works when "leaving" is clearly defined, every score is explained, actions match the reasons, and the model is checked against reality every month.
Try the customer churn prediction demo, see the gym management software, or talk to us about your customers.