How can you use decision trees in the analysis of customer churn?

Official answer

Define churn and the prediction horizon before fitting a tree. For example, predict whether an active subscriber will cancel within the next 30 days, using information available at today's prediction time. Exclude future cancellation records and features created after the outcome.

Build training and evaluation cohorts with complete outcome follow-up. Use temporal evaluation and keep related customer records together where needed. Tune depth, minimum leaf size and pruning on training/validation data, then assess held-out ranking, calibration and performance at the actual contact capacity.

A decision tree splits customers by feature thresholds and assigns a prediction from the resulting leaf. The paths can explain the model's rule, but they do not establish why a customer churns. Correlated features, selection and unmeasured factors can all affect the pattern.

Compare against a simple baseline and inspect performance for important groups. If the goal is retention, test the intervention separately: high churn risk does not necessarily mean a customer benefits from a particular offer. A randomized campaign can measure incremental retention and cost.

Source

scikit-learn: decision trees.