What is a churn risk score?

A churn risk score is a number attached to each customer that says how likely they are to stop paying soon. Its job is not to be a precise probability. Its job is to put the customers in order, so that the first ones on the list are the ones worth a conversation today.

What goes into it

For a subscription business the inputs come from three places:

• Billing. Scheduled cancellations, failed payments, expiring cards, shrinking invoices, refunds and disputes, trials and discounts about to end, renewals coming up, how long the customer has been paying. Every subscription business has this data, because the billing provider keeps it.
• Product usage. Logins, active users, key features used, and how all of those are trending. Strong signals, but only if you send the events somewhere that can read them.
• Relationship. Support tickets, NPS, emails unanswered, a champion who left. Valuable, and the hardest to get into a system.

A good score starts from billing because it is universal and already there, adds usage when available, and leaves relationship data to the human reading the list.

Rules or a model?

Rules give points for each signal and add them up. “Scheduled cancellation: 50 points. Card expiring in 45 days: 20.” Anyone can read why an account scored what it did, the rules can be changed in a minute, and they work from the first day with no training data.

A model learns weights from historical churn. It can find interactions a person wouldn’t write down, and on a large customer base with rich usage data it can rank better. It also needs enough past churn to learn from (hundreds of examples, not dozens), it drifts as the product changes, and its output is a probability with no reason attached, which makes the list hard to act on.

For most companies under a few thousand customers, rules win: they are explainable, they are checkable, and the lift from a model over well-chosen rules is smaller than the cost of building and trusting it. A sensible middle ground is rules whose weights you tune against your own history.

What makes a score trustworthy

1. A reason next to the number. “72, high risk” is useless on its own. “72: set to cancel on 4 December, portal feedback ‘too expensive’” is a task.
2. A backtest. Replay the scoring over the customers who cancelled in the last year. How many did it flag, how many days before the end, and how many accounts did it flag that stayed? If nobody can answer that, the score hasn’t been tested.
3. Stable ordering. A score that reshuffles the list every day for no visible reason trains people to ignore it. The score should move when something happens to the account.
4. A view of how it changed. The customer at 40 who was at 10 last week is more interesting than the one who has been at 45 for a year. A timeline per account, and a “getting worse” view, are worth more than the absolute number.
5. A loop back to outcomes. Record what you did about each flagged account and whether they stayed. Revenue saved out of revenue flagged is the number that tells you whether the score, and your response to it, are worth anything.

Thresholds

Pick two lines: one above which an account is high risk and should be contacted this week, one for medium, worth a look. Set them so the high list is short enough to actually work through. If the high list has sixty accounts and you can call five, the threshold is too low or the rules too loose.

How Churnahead does it

Churnahead’s score is rules-based and capped at 100, with 50 or more as high risk and 25 or more as medium. The rules and their points are published in the documentation, every account shows its reasons in plain words, and the rules are replayed on your last twelve months of churn after every sync so you can see what they would have caught. You can mute any rule or write your own, and each account keeps a 90-day timeline of its score. It reads billing data with read-only access and, if you send events, product usage too.