A practical guide to customer churn prediction: which models actually work, which features move the needle, and how to convert prediction into incremental save rate.
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 comparisonsCustomer churn prediction is the practice of estimating, for each customer, the probability of ending their relationship with the business within a defined horizon. In subscription businesses, that horizon is usually the next billing cycle. In transactional B2C, it is a lapse window (30, 60, 90 days without purchase). In both, the goal is the same: identify who is likely to leave in time to do something about it.
That framing is where most churn programs go wrong. A churn score is a probability, not a program. Two customers with the same predicted churn probability can have completely different treatment-effect economics, and the retention program that treats them identically will over-discount one and under-serve the other. This guide covers both halves, the model that predicts, and the decisioning layer that turns prediction into incremental save rate.
A well-specified churn model estimates:
Freeze features at time t. Do not leak information from after t into the feature set. This is the single most common source of models that look brilliant in cross-validation and fail in production. Any feature that includes post-prediction behavior, including anything that touches the churn event itself, inflates AUC by 5 to 20 points and makes the model useless downstream.
On tabular customer data, gradient-boosted trees (XGBoost, LightGBM, CatBoost) reliably outperform logistic regression, random forests, and most deep-learning architectures. They handle categorical variables cleanly, are robust to missing data, and require little tuning to reach 90 percent of their production-quality ceiling.
Deep-learning architectures start to matter only when sequence structure dominates the signal, long behavioral traces in streaming or gaming, where the ordering of events carries more information than their counts. Even then, boosted trees on well-engineered sequence features usually match transformers within a percent or two of AUC and are ten times cheaper to operate.
Propensity tells you who is likely to churn. It does not tell you whether an intervention will change that. The model family that answers the second question is uplift, and the practical defaults are meta-learners: T-learner (two separate models on treated and untreated), X-learner (better under imbalance), and R-learner (Neyman-orthogonal, more robust). All of them wrap the propensity workhorse (usually a boosted tree) into a causal frame.
The retention program you actually want ranks by uplift, not by propensity. A customer with 60 percent churn probability and near-zero uplift to any intervention is not who you send a save offer to; a customer with 25 percent churn probability and high uplift is.
Model choice matters less than feature engineering. The four feature families that show up as most important in almost every churn model we see:
Marketing exposure features usually add little to churn prediction unless the base rate of contact is unusually low. The reason: for most customers, message frequency is high enough that variation collapses. Where marketing does help is in the uplift model, as a treatment-side feature explaining why response varies by exposure history.
Not by AUC. AUC ranks customers correctly by risk but says nothing about whether your interventions actually save any of them. The metric that matters is incremental save rate: the difference in retention between treated customers (those whose predicted risk triggered an intervention) and a preserved holdout of equally-scored customers who received no intervention.
Reported honestly, this looks like: "Among customers in the top risk decile who received the Save Offer, retention was 62 percent. Among the holdout of top-risk-decile customers who received nothing, retention was 54 percent. Incremental save rate is 8 percentage points." Multiply by predicted margin, subtract intervention cost, report to the CFO.
For a deeper look at how incremental save rate compounds in subscription categories (and where it doesn't), see the economics of churn prevention in streaming and telecom.
A production-grade churn program has three coupled layers:
This is exactly the shape of Next Best Action, applied specifically to the retention lever. Anything less than the full loop leaves incremental save rate on the table.
If you have prediction but not decisioning, start with a preserved holdout and one uplift-ranked retention action. If you have neither, start with a boosted-tree propensity model on the four feature families above; the uplift model can come in the next sprint.
Markin runs churn prediction and uplift-ranked retention as one loop against your base. Explore Product diagnosis to see it live.
