Customer churn prediction
- Explored 14,606 SME customers and 193,002 monthly price records.
- Engineered 20 features: tenure, contract timing, December to January price changes, peak and off-peak spreads.
- Trained a 500-tree random forest and weighted the rare churn class 1:6 so the model stopped ignoring it.
- Finding for the executive summary: price sensitivity was a weak churn driver. I recommended targeting retention on consumption and margin signals instead.