Telecom Churn Prediction

The problem

A churn model tuned for accuracy tells a retention team nothing actionable.

Approach

Feature engineering on usage and contract data, evaluated with precision, recall and ROC-AUC optimised for retention strategy rather than headline accuracy. Interactive Streamlit interface returns churn risk with interpretable driver explanations.

How it works

• Feature engineering over usage and contract data, reaching 80.4% accuracy.
• Evaluated with precision, recall and ROC-AUC, tuned for the retention decision rather than headline accuracy.
• A Streamlit interface returns a churn risk with its drivers. Contract type dominates: month-to-month churns at 42.7% against 2.8% on two-year contracts.

Key decisions

Optimise for the retention decision

Retention budgets are finite, so the cost of a false positive and a false negative are not equal. The threshold is set against that trade-off rather than against accuracy.

Return the driver, not just the score

A retention team can act on 'contract type and support calls' in a way it cannot act on '0.78'.

What the measurements showed

• Accuracy 80.4%, ROC-AUC 0.84, precision 65% and recall 57% on the churn class.
• Contract type dominates: month-to-month churns at 42.7% against 2.8% on two-year contracts.
• Fibre optic customers churn at 41.9%, and customers in their first six months carry the highest risk.

Running it

Full setup, configuration and API reference are in the repository README.

Stack

• Python
• scikit-learn
• Streamlit
• pandas