Retention Risk Analysis

Overview

This project analyzes telecom customer churn from both customer-risk and revenue-risk angles. The goal is to identify which customers are more likely to leave, understand the major churn drivers, and estimate which customer groups can create higher monthly revenue exposure.

Interactive Power BI Dashboard

This dashboard visualizes churn patterns, customer risk levels, and revenue at risk using insights from the ML model. It highlights key KPIs, contract-wise losses, churn probability distribution, and high-value high-risk customers to support data-driven retention decisions.

Approach & Tech Stack

Project Flow: Data Preparation → Churn Analysis → Model Building → Model Evaluation → Risk Scoring → Revenue Impact → Power BI Dashboard

Tools Used: Python, Pandas, NumPy, Scikit-learn, XGBoost, Power BI

Models Evaluated: Logistic Regression, Decision Tree, XGBoost

Modeling & Evaluation

Built and compared three classification models to predict customer churn: Logistic Regression, Decision Tree, and XGBoost.

Key features used in the model included contract type, tenure, monthly charges, total charges, payment method, internet service, and service usage.

Since the primary objective was to identify as many potential churners as possible, recall was treated as the primary model selection metric. This was important because a false negative represents a customer who is likely to churn but is not identified for potential retention action.

Logistic Regression achieved the highest recall at 82.57%, compared with 79.09% for Decision Tree and 70.24% for XGBoost at the selected classification threshold. Therefore, Logistic Regression was selected as the final model.

Although XGBoost achieved higher precision, accuracy, and F1-score, its lower recall meant that it missed more actual churners. The model comparison was therefore driven by the business objective rather than model complexity alone.

Model Performance

Model	Accuracy	Precision	Recall	F1-Score	ROC-AUC
Logistic Regression	0.750	0.518	0.826	0.636	0.862
Decision Tree	0.745	0.512	0.791	0.622	0.839
XGBoost	0.812	0.630	0.702	0.664	0.863

Logistic Regression identified 308 of 373 actual churners, resulting in 65 false negatives. This was lower than the 78 false negatives from Decision Tree and 111 from XGBoost at the evaluated thresholds.

The selected Logistic Regression model was then used to generate churn probabilities, identify high-risk customers (>0.6), and estimate revenue at risk.

Key Findings

Who is Leaving

• About 26.54% of customers are leaving.
• Senior citizens and month-to-month contract customers leave more often.
• Customers paying by electronic check are more likely to leave.
• Churn happens mostly in the first few months of their subscription.
• Customers with 0–3 services leave more often, while customers with 6 or more services usually stay.

Revenue Exposure

The model was used to estimate risk-weighted monthly revenue exposure.

Risk-weighted exposure combines monthly charges with churn probability. It does not mean the company will definitely lose this amount. It shows where revenue is more exposed to churn risk.

• Estimated risk-weighted monthly revenue exposure is around ₹2,13,073.
• High-value customers contribute around ₹1,51,890 of this exposure.
• Month-to-month contracts have the highest churn risk and expected exposure.

Factors Affecting Churn

• More likely to leave: high monthly bills, high total charges, using internet services, not having online backup.
• Less likely to leave: long-term contracts, tech support, phone service, longer tenure, online security.

Recommendations

• Focus on month-to-month customers to reduce the largest possible loss.
• Pay attention to high-value customers with high churn risk.
• Encourage customers to use more services like online backup and tech support.
• Offer discounts or incentives to customers with high bills.
• Engage customers early in their first months to prevent churn.

Impact

• Compared Logistic Regression, Decision Tree, and XGBoost for churn prediction.
• Selected Logistic Regression based on its 82.57% recall and lower number of missed churners.
• Connected churn probability with monthly revenue to estimate revenue exposure.
• Identified customer groups with higher churn risk.
• Created a high-value high-risk customer view for retention priority.
• Supported the Power BI dashboard with model-based churn risk and revenue impact.

Project by Anurag Chauhan