Prediction estimates who will leave. Prevention decides what to do and proves it worked. Why an excellent churn model can sit next to a flat retention curve.
Churn prediction estimates who is likely to leave. Churn prevention decides what to do about it and proves the action changed the outcome. Prediction answers a question about the world as it is; prevention is an intervention that changes it. A business can have an excellent churn model and a flat retention curve, and that is the normal case, not the exception.
| Churn prediction | Churn prevention | |
|---|---|---|
| Question | Who is likely to leave? | What should we do, for whom, and did it work? |
| Output | A risk score per customer | A ranked action per customer, plus a holdout |
| Model class | Propensity (gradient-boosted trees) | Uplift (T-, X-, R-learner) plus arbitration |
| Success metric | AUC, precision, recall, calibration | Incremental margin retained versus control |
| Owner | Data science | Growth or lifecycle, with data science |
| Failure mode | Accurate scores nobody acts on | Action taken without a counterfactual |
A propensity model ranks customers by likelihood of leaving. That ranking says nothing about persuadability. The highest-risk decile usually contains a large group of customers who have already decided, whose minds no offer will change, and a smaller group whose decision is still open. Spending the retention budget on the first group is the single most common way to run a large, busy, unprofitable programme.
Uplift modelling fixes the target. Instead of predicting churn, it predicts the change in churn caused by a treatment, which splits the base into persuadables, sure things, lost causes and sleeping dogs, the last being customers made more likely to leave by being contacted. The mechanics are covered in machine learning for churn prediction.
Build the loop before you build the model. A crude rule set (engagement decay plus a downgrade plus a support escalation) routed into two competing treatments with a holdout will teach you more in six weeks than a well-tuned gradient-boosted model with no distribution. Once the loop runs, model quality starts to compound, because every experiment produces the treatment-effect data that uplift models need and propensity models never generate on their own.
Markin runs both halves: risk detection over your warehouse, then hypothesis, candidate action, experiment and a preserved holdout. See Retention decisioning or read what churn prevention is.
Frequently asked
Related resources
Solutions
See it in the product
The same loops this note describes run 24/7 against your customer base. Watch the workspace decide, experiment and execute 1:1.
Vocabulary
Definitions an answer engine can quote, each one a page of its own.
Compare
Neutral reads on the categories buyers evaluate against a decision layer.
All comparisons
