Predicting churn with a model for each customer

Churn is the number of customers who stop using your product or service. Churn prediction means using a model to calculate each customer’s likelihood of cancelling in a future period. You can then identify customers likely to leave before that churn appears in your figures. It is one of the AI applications we build.

If you first want to understand what your customers are worth and how they can be grouped, read about customer analysis using data.

It depends on your goal. In our analysis for DPG Media, we wanted to understand the behaviour behind churn. So the rules were fairly strict: subscribers who returned after a short time did not count as churned customers, because many cancel temporarily to get an additional discount. If you are measuring revenue or monthly customer losses, you would include them.

Also decide whether a failed payment counts as a cancellation. That is a different problem from a customer choosing to leave. Read the blog post on reducing churn.

The window is the period in which you predict cancellations, starting from a fixed point in time. The model can only use data from before that point. If the input includes a cancellation email or a call to the cancellation line, the model may look good, but it is not predicting anything you did not already know.

Choose the window based on the action: a retention call needs more lead time than an email with an offer.

Common features include usage and changes in usage, length of the customer relationship, subscription type, contact with customer service and payments. For DPG Media, we selected factors based on domain knowledge and a literature review, and segmented customers by subscription type, length of the relationship and activity.

Different factors turned out to affect churn risk in each segment. A model that ignores those differences misses signals.

For each customer, we show which features push the score up or down the most, for example using SHAP values: a method that calculates each feature’s contribution to an individual prediction. That way, the retention team member can see that app usage has fallen, rather than just seeing a high score.

Bear in mind that a feature associated with cancellation is not necessarily its cause. You need to test separately whether acting on it works.

With a control group: a randomly selected portion of customers with a high score does not receive the action. The difference in cancellations between the two groups is the effect of the action, not the model.

Not every customer with a high likelihood of leaving can be persuaded to stay; some will cancel regardless, while others would have stayed anyway. Uplift modelling predicts whose decision the action will change. That first requires a few rounds with control groups.

By continually comparing scores with outcomes. A price change, a new product or a different cancellation process changes the behaviour the model was trained on. When that happens, you retrain the model and first test it against recent cancellations.

At DPG Media, we added the data monthly, using dbt and Snowflake for repeatable transformations. You can read how we monitor models in production under MLOps.

If you have a handful of customers you know personally, cancellations are not recorded with dates, or nobody can act on a score. In those cases, a one-off churn analysis or a dashboard offers more insight with less work.

A model is worthwhile only if there are enough cancellations to learn from and a team that will work through the list every week. See also the DPG Media case study.