🏦 Bank Customer Churn Prediction & Risk Scoring

A machine learning project that predicts the probability of bank customer churn and provides an interactive risk-scoring dashboard using Streamlit.

📌 Project Overview

Customer churn can affect revenue stability, customer lifetime value, and long-term banking relationships. This project uses historical customer data to identify patterns associated with churn and build a model that estimates the probability of a customer leaving.

The project was developed as part of a Unified Mentor internship project.

🎯 Objectives

• Predict whether a bank customer is likely to churn.
• Generate a customer-level churn probability.
• Identify the most influential churn-related features.
• Compare multiple machine learning models.
• Provide an interactive dashboard for churn risk assessment.
• Demonstrate simple what-if analysis for customer engagement and product usage.

📊 Dataset

The dataset contains 10,000 European bank customer records.

Target Variable

• Exited = 0 → Retained customer
• Exited = 1 → Churned customer

Main Features

• CreditScore
• Geography
• Gender
• Age
• Tenure
• Balance
• NumOfProducts
• HasCrCard
• IsActiveMember
• EstimatedSalary

CustomerId and Surname were excluded because they are identifiers. Year was also excluded because it did not provide meaningful variation for the model.

Churn Distribution

• Retained customers: 7,963
• Churned customers: 2,037
• Overall churn rate: 20.37%

🔧 Feature Engineering

Four additional features were created:

1. Balance-to-Salary Ratio

   • Balance / EstimatedSalary

2. Product Density

   • NumOfProducts / (Tenure + 1)

3. Engagement × Product Interaction

   • IsActiveMember × NumOfProducts

4. Age × Tenure Interaction

   • Age × Tenure

🤖 Machine Learning Models

Three classification models were compared:

• Logistic Regression
• Random Forest
• Gradient Boosting

The dataset was split using an 80:20 stratified train-test split.

Model Performance

Model	Accuracy	Precision	Recall	F1 Score	ROC-AUC
Gradient Boosting	87.05%	79.13%	49.39%	60.82%	86.74%
Random Forest	86.50%	77.51%	47.42%	58.84%	85.41%
Logistic Regression	80.60%	57.04%	18.92%	28.41%	77.35%

Gradient Boosting was selected as the final model because it achieved the strongest overall performance.

📈 Feature Importance

The most influential features in the final Gradient Boosting model were:

Rank	Feature	Importance
1	Age	0.387126
2	Number of Products	0.285662
3	Engagement × Product Interaction	0.067163
4	Balance	0.063198
5	IsActiveMember	0.058565
6	Geography: Germany	0.053754
7	Balance-to-Salary Ratio	0.023583
8	Credit Score	0.017604
9	Estimated Salary	0.015288
10	Gender: Male	0.010199

These values represent the model's relative feature importance and should not be interpreted as causal effects.

🔎 Key EDA Findings

• Overall churn rate was 20.37%.
• Germany had the highest observed churn rate at 32.44%.
• Spain had a churn rate of 16.67%.
• France had a churn rate of 16.15%.
• Inactive members had a churn rate of 26.85%, compared with 14.27% for active members.
• Churned customers were older on average: 44.84 years versus 37.4