One-line pitch: A full-stack business analytics project that translates raw telecom customer data into a calibrated predictive model, surfaces the drivers of churn using advanced SQL and SHAP, and calculates the exact ROI of a targeted retention campaign.
Customer churn is the silent killer of subscription businesses. This project goes beyond simply training a machine learning model; it answers the "so what?" for business stakeholders. By combining predictive modeling with financial ROI simulation, we identify exactly who is going to churn, why they are leaving, and exactly how much money we can save by intervening.
1. The Retention ROI Pipeline Using a calibrated Random Forest model, we scored all 7,043 active customers and calculated the Expected Value of Action (EVA) assuming a $50 retention offer and a 30% intervention success rate. We identified 4,074 high-risk customers where an intervention yields a positive ROI, resulting in $744,395 in projected net savings over a 24-month horizon.
2. The Contract Trap (SQL & SHAP Drivers) Both our advanced SQL cross-tabs and our SHAP Machine Learning explainers independently isolated Contract Type and Tenure as the massive, overwhelming drivers of churn. Customers on Month-to-Month contracts in their first 10 months are the highest flight risk.
The business should immediately deploy the $50 retention campaign to the top 4,074 ranked customers identified in the retention_targets.csv file, specifically targeting Month-to-Month customers with high monthly charges to lock them into 1-year contracts.
graph TD;
CSV[Telco Raw Data] -->|SQL & Pandas| EDA[Driver Analysis & Segmentation];
CSV -->|Scikit-Learn| ML[Calibrated Random Forest];
ML -->|SHAP| EXPLAIN[Feature Explainability Plot];
ML -->|Probabilities| ROI[Retention ROI Simulator];
ROI --> TARGETS[retention_targets.csv];
EDA --> PBI[Power BI Dashboard];
TARGETS --> PBI;
classDef raw fill:#e1f5fe,stroke:#0288d1;
classDef model fill:#cd7f32,stroke:#8b5a2b;
classDef output fill:#ffd700,stroke:#daa520;
class CSV raw;
class ML,EDA,ROI model;
class PBI,TARGETS,EXPLAIN output;
The 3-tab Streamlit dashboard allows non-technical stakeholders to slice the risk pools, view live SHAP explainability models, and interactively adjust ROI assumptions to instantly download the targeted intervention lists.
We prioritized AUC-PR over Accuracy because churn is an imbalanced class. Predicting "no one will churn" yields an artificially high accuracy of 74%.
We wrapped our Random Forest in a CalibratedClassifierCV to ensure the probabilities output by predict_proba() were mathematically sound enough to use in financial calculations.
- AUC-PR: 0.659
- Precision: 0.68
- Recall: 0.46
- F1 Score: 0.54
We used SHAP (SHapley Additive exPlanations) to crack open the "black box" of the Random Forest.
To convert model probabilities into a business decision, we built a scoring table that ranks customers by the expected value of intervening.
The Financial Logic:
CLV = MonthlyCharges * 24 (months)
Expected_Loss = Churn_Probability * CLV
Expected_Value_of_Action = (Expected_Loss * 30% Success Rate) - $50 Intervention CostBy targeting only customers where the Expected_Value_of_Action > 0, we ensure the retention campaign is strictly profitable.
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txtRun the scripts sequentially to generate the calibrated model, SHAP plots, and financial targets:
cd src
python 1_data_cleaning.py
python 3_model.py
python 4_retention_roi.pyThe 5 advanced analytical queries (window functions, NTILE, cross-tabs) used to validate the business drivers are available in the /sql directory. Run them against the raw data using DuckDB or SQLite.
Open the /powerbi/ folder and load the .pbix file to interact with the dashboards.