Automation and AI

Churn Prediction

Also called churn modelling, churn modeling

Using past customer behaviour to estimate which current customers are likely to cancel or stop buying, so you can act before they leave.

Quick facts: Churn Prediction

Category
Automation and AI
Also called
churn modelling, churn modeling
Level
Intermediate
Affects
Retention, customer lifetime value, discount spend, email and CRM programmes
Where to see it
CRM risk scores, email platform predictive fields, spreadsheets, Python or BigQuery for custom models
In this article4
  1. How churn prediction works
  2. Why it matters
  3. Common mistakes
  4. How to act on it

Churn prediction is the practice of using what you know about past customers to estimate which current customers are likely to cancel, lapse or stop buying in the near future. Instead of counting losses after the event, it gives each customer a risk score so you can decide who to contact, what to offer and when.

How churn prediction works

The starting point is a clear definition of churn for your business. For a subscription box, it is a cancelled plan. For a gym, it might be a membership that is frozen and then not restarted. For a shop with no contracts, churn has to be inferred: a customer who usually orders every six weeks and has not ordered for four months has probably gone. Your churn rate measures how many leave; churn prediction tries to say which ones.

Once churn is defined, you look back at customers who left and those who stayed, and compare what they did beforehand. Typical signals include:

  • Falling engagement Fewer logins, unopened emails, shorter sessions in the app.
  • Changing purchase patterns Longer gaps between orders, smaller baskets, a switch to discounted items only.
  • Service friction Complaints, refunds, failed payments or a support ticket that took a week to close.
  • Account changes A downgrade, a removed card, or a cancelled direct debit.

The simplest version is a set of rules, such as “no order in 90 days and last email unopened”. A step up is RFM analysis, which ranks customers by recency, frequency and spend. The most advanced version is a propensity model built with machine learning, which weighs dozens of signals at once and outputs a probability for each customer. Many CRM and email platforms now include a built-in churn or “at risk” score, though how they calculate it is rarely explained in detail.

Why it matters

Keeping a customer usually costs less than winning a new one, and a customer who stays longer is worth more over their lifetime. Churn prediction connects those two facts to action. If you know which customers are drifting, you can put your retention budget where it changes outcomes rather than spreading discounts across everyone.

It also protects margin. A blanket “we miss you” voucher goes to people who were going to buy anyway. A targeted offer to the customers most at risk costs less and tells you whether the offer actually changed anything.

For UK subscription businesses there is a regulatory angle as well. The Digital Markets, Competition and Consumers Act 2024 brings in rules on subscription contracts, including reminders and easy cancellation, as those provisions come into force. Retention work has to sit inside those rules: the aim is to fix the reason someone is leaving, not to make leaving harder.

Common mistakes

  • A vague definition of churn. If the team cannot agree when a customer counts as lost, the model has nothing reliable to learn from.
  • Predicting without a plan. A risk score that nobody acts on is a report, not a programme. Decide the action for each risk band before building the score.
  • Measuring the wrong thing. A save campaign looks successful if you only count customers who stayed. Without a holdout group that receives nothing, you cannot tell how many would have stayed anyway.
  • Using only marketing data. Billing failures and support history often predict churn better than email opens, especially since Apple Mail Privacy Protection inflated open rates.
  • Treating everyone at risk the same. A loyal customer having a bad month needs a different message from a bargain hunter who only ever bought on promotion.

How to act on it

Start with a spreadsheet before you reach for software. Export customers with their last order date, order count, total spend and any complaint or refund flag. Look at who left last year and see which of those fields separated them from the people who stayed. That simple exercise often shows one or two signals that matter far more than the rest.

Next, set up three bands (low, medium and high risk) and agree an action for each: perhaps nothing for low risk, a useful content email for medium, and a personal check-in or service fix for high. Hold back a small random group from each band so you can measure the real effect.

Finally, compare what a saved customer is worth with what the save costs. If the numbers do not work, change the offer rather than the model. Building this into a wider retention and acquisition plan is part of the digital marketing strategy work I do, where churn sits alongside acquisition cost and customer lifetime value.

Do and do not

Do

  • Agree a precise definition of churn first
  • Decide the action for each risk band before you build the score
  • Keep a holdout group to measure the real effect

Do not

  • Send the same discount to every at-risk customer
  • Rely on email opens alone as a signal
  • Make cancelling harder as a retention tactic

Questions people ask about this

How much data do I need for churn prediction?

For a rules-based approach, a few hundred customers with a year of history is enough to spot patterns. A machine learning model needs more, ideally thousands of customers and a good number of people who actually churned, because it learns from those examples. If you have a small customer base, RFM scoring and simple rules usually beat a model trained on too little data.

Is churn prediction only for subscription businesses?

No. It is easiest for subscriptions because cancellation is a clear event, but shops, salons, trades and B2B suppliers all lose customers quietly. In those cases you define churn as a gap longer than a customer's normal buying cycle, and predict who is approaching that gap.

Can I use customer data for churn prediction under UK GDPR?

Usually yes, if you have a lawful basis such as legitimate interests, you tell customers in your privacy notice that you analyse their behaviour to improve service and retention, and you keep the data proportionate. If the score leads to decisions with legal or similarly significant effects on someone, the rules on automated decision-making need a closer look, so take advice in that case.

Related terms

Found this useful?

Share it, or ask an AI to summarise it

Back to the glossary

Knowing the term is the easy part

Applying it to your own site and budget is the work. Book a call and I will tell you what actually applies to you.