Customer churn prediction
Who is leaving, with how much lead time and why
The problem
Companies with recurring customers have a risk formula. Four or five signals, a weight for each, a threshold that separates at-risk customers from the rest. Someone built it with good judgement and for a while it worked.
Then the business changed and the formula did not. It treats a three-month customer and a four-year customer the same, does not distinguish between someone who is leaving and someone going through a rough patch, and its weights are still the ones someone chose back then. The risk list ends up full of people who were not going to leave, and those who actually do don't appear. The team executes well on a badly assembled list.
It is an expensive problem at a time when almost no one can afford it.
In Gartner's 2025 survey of revenue leaders, 73% declared they prioritise growth over their existing customer base. Acquisition costs have risen steadily, and conversion on cold audiences does not compensate. The risk list stopped being an internal exercise and became the input for most marketing decisions.
The solution
A model that calculates, for each customer, the probability of cancelling within a defined time window, and the signals that carry the most weight in that specific case.
The reason is used as much as the probability. A customer at risk due to price and another due to a drop in activity need different actions, and the reason is what routes each one. The list arrives sorted, so it is trimmed by budget rather than by threshold. If this week there is capacity for three hundred contacts, those are the first three hundred.
What the model does not do is decide who to contact. That depends on each customer's margin, the cost of the action, and the team's capacity. Those are things your team knows and the model does not, and they can be added later as a decision layer on top of the model.
Why a model and not a better formula
Because the problem is not that the formula has too few signals. It is that the signals interact. A drop in frequency does not carry the same weight if it came after a price increase, and that combined effect changes by segment and by point in the cycle. A weighted sum cannot capture that without becoming unmanageable.
And because a fixed formula ages. The model is retrained with the history that keeps accumulating, learns new patterns as they appear, and allows quick adjustments when a new data source becomes available.
More on the blog: Before training a churn model, there is a business question that determines the metric to optimise →
How it integrates
The result is delivered where the team already works, whether the CRM, the campaign tool, or the data warehouse. Each prediction comes with the signals that explain it, which is what allows choosing the action and, in regulated sectors, justifying why a customer appeared on a list.
The model is kept under watch.
A study by researchers at MIT, Harvard, and the University of Monterrey found that 91% of machine learning models degrade over time. It is not a construction defect: it is what happens when the world the model learned from stops resembling the one in front of it.
That is why both data drift and accuracy are measured continuously, and retraining is triggered before the decline reaches the results.
The same model, other applications
A churn model predicts the end of a recurring relationship from deterioration signals. The data changes, not the structure.
Employee turnover.
Which people are most likely to resign in the coming months. Consumed by HR and team managers.
Default and delinquency.
The same mechanism applied to a credit relationship, feeding collections policy instead of campaigns.
B2B account loss.
Reduction in orders, drop in contacts, change of counterpart. Consumed by the account sales team.
Programme dropout.
Adherence to a treatment, retention in a training programme. Consumed by the follow-up team.