Back to Projects
PREDICTIVE ANALYTICS$2.4M ANNUAL SAVINGS

Customer Churn Prediction

Production ML system predicting customer churn with 89% accuracy, enabling proactive retention campaigns that reduced churn rate by 23% and saved $2.4M annually in customer lifetime value.

Model Accuracy

89%

Annual Savings

$2.4M

Churn Reduction

23%

Customers Analyzed

500K+

Random ForestXGBoostSMOTEFeature EngineeringA/B TestingPython

Business Impact

The Problem

The company was losing 15% of customers annually with no early warning system. Marketing spent $500K+ on broad retention campaigns with only 8% response rate, wasting resources on customers unlikely to churn while missing high-risk segments.

  • • 15% annual churn rate ($10M+ lost revenue)
  • • No predictive capabilities for at-risk customers
  • • Untargeted retention campaigns (8% response)

The Solution

Built an ML pipeline that scores all 500K+ customers daily, identifying high-risk segments 30 days before likely churn. Enabled targeted retention with personalized offers, increasing campaign effectiveness by 3x.

  • • 89% accuracy in predicting churn 30 days out
  • • 23% reduction in overall churn rate
  • • 3x improvement in campaign response rate

Interactive Demo

Try the model yourself - enter customer data to see real-time churn predictions

Customer Information
Enter customer details to predict churn probability

Enter customer information to see prediction results

Model Performance Metrics
Evaluation on test dataset (1,409 samples)

Technical Implementation

Model Architecture

  • • Stacked ensemble: Random Forest + XGBoost + LightGBM
  • • Meta-learner for final probability calibration
  • • SMOTE for handling class imbalance (73/27 split)
  • • Bayesian hyperparameter optimization
  • • 5-fold stratified cross-validation

Feature Engineering

  • • RFM analysis (Recency, Frequency, Monetary)
  • • Customer lifetime value predictions
  • • Behavioral cohort segmentation
  • • 45+ engineered features from raw data
  • • Automated feature selection via Boruta

Production Pipeline

  • • Daily batch scoring of 500K+ customers
  • • Real-time API for individual predictions
  • • Model monitoring with drift detection
  • • A/B tested retention campaigns
  • • Automated retraining on new data

A/B Testing & Experimentation

Validated the model's business impact through rigorous A/B testing before full deployment:

Test Duration

8 weeks

Sample Size

50,000 customers

Statistical Significance

p < 0.001

Result: Treatment group (ML-targeted retention) showed 23% lower churn vs. control (traditional targeting), with 95% CI [19%, 27%]. Lift in campaign response rate: +215%.