Telecom customer churn prediction

Identifying high-risk customers before they cancel, enabling targeted retention campaigns that maximize ROI across a 5,000-customer base.

LightGBM SHAP Classification Cost-benefit analysis Telecom Customer retention

$141K

Estimated annual savings

Interactive dashboard

Five-page Streamlit application for exploring churn patterns and model outputs

Overview

Dataset summary and churn distribution
Key risk factors at a glance
Model selection and best parameters

Exploratory analysis

Churn rate by contract type and tenure
Monthly charges distribution by segment
Correlation heatmap of service features

Model performance

Four-model comparison with ROC curves
Confusion matrices and classification reports
Threshold optimization analysis

Explainability

SHAP summary and waterfall plots
Global and local feature importance
Individual customer prediction breakdown

Business impact

Cost-benefit curve across classification thresholds
Optimal threshold for net savings with $50 intervention cost vs. $75/month retained revenue
Full-base annual savings projection at 30% retention success rate

Launch live app Read case study

$ pip install -r requirements.txt && streamlit run app.py

Key results

LightGBM selected as the best model after four-model comparison with GridSearchCV

0.713

AUC-ROC

85%

Recall (churners caught)

$141K

Annual savings estimate

5,000

Customers analyzed

Methodology

Built on 5,000 telecom customer records with 20 features covering demographics, service subscriptions, billing, and tenure. Four classifiers were trained with GridSearchCV hyperparameter tuning. SHAP values provide both global and individual-level explanations. A cost-benefit framework with $50 intervention cost and $75/month retained revenue quantifies the business case for targeted retention.

Data preparation

Imputation, encoding, feature engineering

Model training

LR, RF, XGBoost, LightGBM

Explainability

SHAP global and local

Business impact

Threshold optimization, cost-benefit