← All posts

· ai, machine-learning

Predictive churn, explained

How TrackCrumb scores churn risk with XGBoost and explains every prediction with per-user SHAP attribution.

Knowing that a user churned is history. Knowing who is about to churn — and why — is leverage. TrackCrumb's predictive churn turns your event stream into a nightly risk score for every user and account, available on the Scale tier and up.

How the model works

  1. Features — we derive behavioural features from your events: recency, frequency, depth of use, and trend.
  2. Scoring — an XGBoost model produces a churn probability per user, retrained nightly so it tracks your product as it changes.
  3. Survival curves — a Cox proportional-hazards model estimates when risk rises over the next 30 days.

Every score is explainable

A black-box risk number is useless to a growth team. That's why each prediction ships with SHAP attribution — the specific signals pushing a user toward or away from churn. "This account is high-risk because weekly active seats dropped 40% and they stopped using funnels" beats a bare 0.82.

From insight to action

Predictive scores feed directly into campaign targeting, so you can trigger an in-app message or email the moment risk crosses your threshold — no export, no reverse-ETL.

Piyapath

Piyapath

Founder, TrackCrumb

Building TrackCrumb — AI-native product analytics. Writing about analytics, instrumentation, and shipping a solo SaaS.