Goal — Predict churn and understand why customers leave so telecom teams can intervene earlier.
Stack — Python (pandas, numpy, scikit-learn, matplotlib, seaborn), SHAP; Jupyter Notebook.
Data — Telco Customer Churn (Kaggle), ~7,000 customers, 21 attributes including contract, tenure, and billing info.

At a glance

  • Built a complete churn prediction pipeline using supervised learning, with clean preprocessing and model comparison.
  • Trained a Random Forest classifier and benchmarked it against Logistic Regression for interpretability.
  • Used SHAP to explain model decisions — both globally (feature ranking) and locally (case-specific explanations).
  • Extracted business-focused recommendations tailored to retention, pricing strategy, and payment behavior.

Data & preparation

  • Cleaned missing values and converted TotalCharges from string to float for numerical operations.
  • Dropped unique IDs (customerID) and encoded categorical features with one-hot encoding.
  • Scaled continuous columns like MonthlyCharges and Tenure to prep for logistic models.
  • Created visual summaries (correlation heatmaps, barplots) to understand churn patterns before modeling.

Models & evaluation

  • Trained a Random Forest as the primary model due to its ability to handle non-linear interactions.
  • Logistic Regression was used as a baseline model for quick interpretability.
  • Evaluation was done using confusion matrix, classification report, and precision/recall curves.
  • Focused on optimizing recall for the churn class, ensuring fewer at-risk customers were missed.

Why these two?
Logistic Regression offers transparency and coefficients; Random Forest finds complex churn patterns like tenure × contract type. Together they show accuracy vs explainability tradeoffs.

Explainability

  • Applied SHAP (SHapley Additive Explanations) to both models for post-hoc interpretability.
  • Top global features: Contract, Tenure, MonthlyCharges, and PaymentMethod emerged as churn signals.
  • Local SHAP plots helped explain why individual customers were flagged, useful for support teams and marketing segmentation.

Business insights

  • Short-tenure customers on month-to-month contracts had the highest churn — loyalty risk.
  • Higher billing amounts correlated with churn, possibly tied to perceived value or pricing concerns.
  • Customers paying with electronic check churned more than those using credit cards or autopay.
  • Data suggests that billing friction and lack of contract commitment are key retention factors.

Recommendations

  • Incentivize annual contracts for new users (<12 months) to improve loyalty.
  • Add value to high-cost plans through bundled services or promotions.
  • Promote autopay enrollment to reduce churn tied to billing inconvenience.
  • Use the model to trigger outbound save campaigns and early alerts based on churn scores.

Reproduce

  1. Download the Telco Churn dataset from Kaggle.
  2. Open telco_churn_analysis.ipynb in Jupyter Notebook.
  3. Follow the cells in order — from data cleaning to modeling, SHAP, and insight generation.

🔗 View the code on GitHub

Next

  • Add probability calibration (Platt scaling / Isotonic regression) for marketing thresholds.
  • Incorporate XGBoost with monotonic constraints for robust churn modeling.
  • Build an inference script to monitor class balance, model drift, and prediction quality over time.

Updated: