🏦 Customer Churn Prediction for Financial Institutions

A machine learning project that predicts customer churn in financial institutions using various classification algorithms and advanced techniques like SMOTE and ensemble methods.

📊 Project Overview

This project implements multiple machine learning models to predict customer churn in banking institutions. The best-performing model achieves a recall of 0.59 for churned customers, making it particularly valuable for institutions where customer retention is a priority.

Key Features

• Multiple classification algorithms comparison
• Feature engineering for enhanced prediction
• Handling class imbalance using SMOTE
• Ensemble methods implementation
• Comprehensive model evaluation metrics

🛠️ Technologies Used

• Python 3.7+
• pandas
• scikit-learn
• XGBoost
• seaborn
• matplotlib
• imbalanced-learn

📈 Models Implemented

• XGBoost Classifier
• Random Forest
• Decision Tree
• Support Vector Machine (SVM)
• K-Nearest Neighbors (KNN)
• Naive Bayes
• Logistic Regression
• Voting Classifier (Ensemble)

🚀 Getting Started

Prerequisites

pip install pandas numpy scikit-learn xgboost seaborn matplotlib imbalanced-learn

Dataset

The project uses a customer churn dataset with the following features:

• Demographics (Age, Geography, Gender)
• Banking relationships (Balance, Credit Score, Tenure)
• Product usage (NumOfProducts, HasCrCard, IsActiveMember)
• Financial metrics (EstimatedSalary)

Running the Models

# Clone the repository
git clone https://github.com/stonewerner/customer-churn-ML.git
cd customer-churn-ML

📊 Feature Engineering

Several derived features were created to improve model performance:

• Customer Lifetime Value (CLV)
• Age Groups (Young, MiddleAge, Senior, Elderly)
• Tenure-Age Ratio
• One-hot encoded categorical variables

💡 Key Findings

• The ensemble model with SMOTE achieved the best recall (0.59) for churned customers
• Most important features for prediction:
  • Balance
  • Age
  • EstimatedSalary
  • Geography
  • NumOfProducts

📝 Model Selection Rationale

The final model prioritizes recall over precision because:

• The cost of losing a customer (false negative) is typically higher than the cost of retention actions on non-churning customers (false positive)
• Higher recall ensures we identify more potential churners, allowing proactive retention measures

🤝 Contributing

Feel free to fork the repository and submit pull requests. For major changes, please open an issue first to discuss the proposed changes.

👥 Contact

For questions or collaboration opportunities, please reach out to Stone Werner, stonewerner.com

🔍 Future Improvements

• Deep learning models implementation
• API deployment for real-time predictions
• Feature selection optimization
• Hyperparameter tuning
• Cross-validation implementation