Churn prediction: how to spot the customers about to leave, and act in time

A churn prediction model estimates how likely each customer is to stop buying, cancel, or lapse within a defined window, so you can act before they go. To work, it needs three things: behavioural signals that change before a customer leaves (declining usage, slower logins, support tickets, missed payments), a clear definition of what "churned" means and a labelled history of who actually left, and a decision attached to the score so a high-risk flag triggers a real intervention. The model produces a probability per customer. On its own that number does nothing. Reducing churn is a separate discipline: it depends on what you do with the at-risk list, not on how accurately you predicted it. Treat the score as the start of the work, not the end of it.

What signals does a churn prediction model need?

A model can only see what you record. The useful signals are the ones that move before a customer leaves, not after.

• Engagement trend, not just level. A customer dropping from daily to weekly use tells you more than their absolute usage. Slope matters more than position.
• Recency and frequency. Days since last login, last order, or last meaningful action. Lengthening gaps are an early warning.
• Support and friction. Ticket volume, repeated failed actions, complaints, time to resolution.
• Commercial signals. Failed payments, downgrades, declined renewals, falling order value.
• Lifecycle context. Tenure, plan type, onboarding completion. A three-week customer and a three-year customer churn for different reasons.

The honest constraint: if a signal is not captured today, the model cannot use it tomorrow. Often the first real piece of churn work is fixing what you log, not training anything.

How do you label churn so the model can learn?

Labelling is where most churn projects quietly go wrong, because "churn" is not one thing. You have to define it before you model it.

1. Pick the churn event. A cancellation is easy. A quiet customer who simply stops ordering is harder. For non-subscription businesses, you usually define churn as no activity within a rolling window (for example, no purchase in 90 days).
2. Choose the prediction window. Are you predicting churn in the next 30, 60, or 90 days? The window has to match how long your intervention needs to work.
3. Build the history. Take a past point in time, use only the data available then, and label whether each customer churned after it. Using information that did not exist yet (a leakage error) makes a model look brilliant in testing and useless in production.
4. Check the base rate. If only a small share of customers churn, accuracy alone is misleading. A model that predicts "nobody leaves" can score highly and help no one.

What is the difference between predicting churn and reducing it?

This is the distinction that decides whether the project pays for itself.

	Predicting churn	Reducing churn
Output	A risk score per customer	Fewer customers actually leaving
Owned by	Data and engineering	The team that talks to customers
Measured by	Model accuracy, precision, recall	Retention rate, saved revenue
Fails when	Signals are missing or leaked	Nobody acts on the score

A model that ranks your customers by risk is only the diagnosis. Reduction comes from the response: a call from an account manager, a targeted offer, a fix to the onboarding step where people drop off. The score points the response at the right people. It does not make the response happen.

This is the same logic that runs through predictive analytics: turning data into decisions. A forecast nobody acts on is a hobby, however accurate.

How do you act on a churn score without wasting it?

Acting well means being selective, because intervention has a cost and a wrong intervention can do harm.

• Set a threshold tied to capacity. If your team can call fifty customers a week, the score should hand them the fifty most at-risk, not a list of thousands.
• Match the action to the reason. A customer at risk over price needs a different response from one stuck on a broken feature. The signals that drove the score usually tell you which.
• Avoid over-contacting the safe. Reaching out to a happy, low-risk customer with a retention offer can plant the idea of leaving. Precision protects you here.
• Close the loop. Record what you did and whether it worked, then feed that back. The intervention data is what lets the model and the playbook improve.

Done this way, the score becomes a queue your team works through, with the most valuable conversations at the top.

The practical takeaway

Building the model is the smaller half of the job. The larger half is defining churn honestly, capturing the signals that move early, and attaching a real action to the score. If you are scoping a churn prediction model, ask one question before any data is touched: when a customer scores high-risk, what happens next, and who does it? If there is no answer, you have a maths project, not a retention programme. If there is, you have something worth building. This is the kind of work our predictive analytics practice is built around.