The Problem

Retention teams were largely reactive — by the time a customer’s cancellation request came in, there was little left to do. The business needed a way to identify customers likely to churn while there was still time to act.

The Approach

A production churn prediction model was built using PySpark-based ML pipelines, trained on real telecom usage patterns. Key parts of the approach:

• Feature engineering from usage, billing, and support-interaction data
• A model pipeline designed to run on a regular cadence, not as a one-off analysis
• Output wired directly into retention campaign targeting, not left as a static report

The Outcome

Retention teams gained a ranked, regularly-refreshed list of at-risk customers, shifting the workflow from reactive to proactive.

Specific model performance and retention-lift metrics are being prepared for publication and will be added here.

← Back to all work