Machine Learning
What is churn prediction?
Churn prediction is a model that estimates how likely a given customer is to stop using or paying for a product within a defined future window.
The definition problem comes first
Before any modelling, you have to define churn in a way your business agrees on. For a subscription, it is cancellation - easy. For a casino, a marketplace, or a usage-based product, there is no event. You have to pick a window: 'no qualifying activity in the next 30 days' or 'no deposit in the next 45'. Getting this wrong invalidates everything downstream, and it is a business decision, not a data-science one.
How the training data is built
Pick an observation date. Compute features only from data before it. Look forward from that date to the end of the window and label whether the customer churned. Repeat across many historical dates so the model sees seasonality and different lifecycle stages. Evaluate on time-held-out windows - never on a random split, which lets the future leak into training.
What good looks like
Not raw accuracy - on an imbalanced problem a model that predicts 'nobody churns' can score 95%. What matters is precision within the top slice you can actually act on. If your team can contact 200 customers a week, the only question is how many of that ranked top 200 genuinely would have churned. That is the number to optimise and the number to report.
Predicting is not preventing
A perfectly calibrated churn model changes nothing on its own. The value lives entirely in the intervention attached to it, and in whether the intervention reaches the customer while the window is still open. Most churn projects that fail do not fail at modelling - they fail because nobody built the workflow that acts on the output.
We build automation systems for small teams. Free audit call to map your specific workflows - no pitch, just a plan.