Case study 03
Predictive analytics · Python
Turning churn predictions into business decisions.
Can a churn model produce a useful retention strategy instead of another probability score?
- Role
- Analysis, feature engineering, and model evaluation
- Format
- Independent project
- Tools
- Python · Pandas · scikit-learn · LightGBM · PyCaret
- 10,000Customers analysed
- ~20%Churn rate
- 6Models compared
- 0.63Best F1 score
The problem
A churn score only matters if it changes an action.
I analyzed 10,000 retail-banking customers with a churn rate of roughly 20%. The project connected model performance to specific retention actions instead of stopping at a ranked model table.
The business needed enough recall to find at-risk customers, but low precision would waste outreach. No single metric could decide that trade-off without the campaign context.
The approach · 4 steps
Compare models, then connect signals to retention.
- Explored churn patterns across age, geography, balance, activity, and product use.
- Compared six classification models with five-fold cross-validation.
- Created behavioural features such as product usage, engagement, and balance-to-salary ratio.
- Used PyCaret as a second check on the manually built pipeline.
The result
The output was a ranked intervention plan.
- 0.63Random Forest F1
- 0.68Recall on the test set
- 6Models compared
Age, number of products, activity, and balance were the strongest signals. The practical output was a set of targeted retention ideas and a plan for ranking customers by intervention priority.
What I would do next