Features
For each customer, as of a given date:
A feature for a date only uses rows dated on or before it. Status and cancellation columns decide who churned; they are never used as features.
Training and testing
- Snapshots. Every 14 days through your history, each active customer’s features are computed as of that day, and labelled by whether they churned within the prediction window after it (30 days by default).
- Time split. The most recent fifth of snapshot dates is held back for testing. The model is trained on the earlier dates only.
- Model. LightGBM, with early stopping, and class weights for the rarer churners. Scores are calibrated on the held-back weeks with isotonic regression.
- Final fit. The model is retrained on every labelled date, with the same settings, to score today’s customers.
Quality numbers
All measured on the held-back weeks:
Each model gets a verdict:
Reasons
Each score comes with the features pushing that customer’s risk up most, measured with exact Tree SHAP values. Each becomes a sentence from a template, and a sentence is used only when the customer’s own value supports it:
The reasons are written from templates in the Signals service. No AI model writes them.

