Churn: predicting when a customer leaves

Churn modeling is structurally the closest of all seven fields to a generic "time to death" problem — which is exactly why it borrows medicine's validation toolkit wholesale.

The core idea

Customer churn is commonly modeled two different ways: as a binary classification problem (will this customer churn in the next period — yes or no?), scored with AUC/ROC; or as a survival problem (when will this customer churn?), using the exact same Cox proportional hazards or accelerated failure time models introduced on the medicine page, just applied to a subscription cancellation instead of a clinical event. The survival framing is the more informative of the two, because it predicts a moment in time rather than a single yes/no verdict at a fixed horizon.

Benchmarks reported in the literature

Classification-framing AUC, by algorithm

Typical AUROC ranges reported across churn studies
Algorithm	Reported AUROC
Logistic regression	65–72%
Decision trees	60–70%
Random forest	70–78% (one subscription-business study: 0.83)
XGBoost / LightGBM	72–82%
Ensemble-fusion models (best case)	0.91
Rental-service ML model	0.88 AUC, F1 0.93

A commonly cited 2026 industry benchmark describes 70–85% accuracy as typical for "production-grade" churn models in real-world deployment.

Survival-framing (Cox PH) concordance — directly comparable to the C-index bar

Because churn survival models use the identical concordance metric introduced in the medicine article, results are directly comparable across the two fields:

A survival-analysis approach on the IBM Telco churn dataset reported a concordance of 0.867.
A separate telecom Cox-PH churn study reported concordance of 72% for a 6-month product and 79% for a 12-month product — concordance improving as the observed horizon lengthens.

How this field validates its predictions

Survival-framed churn models are validated with the exact same toolkit described on the medicine page: C-index for discrimination, Brier score and calibration slope/intercept for calibration. This is the one field surveyed here where an older discipline's validation math transfers to a business problem with literally zero adaptation — it is the same Cox or AFT model, just fit against a business event (cancellation) instead of a medical one (death or disease progression).