🏦 Customer Churn Prediction using Machine Learning

An end-to-end machine learning project that predicts customer churn in a banking dataset — featuring full EDA, statistical feature selection, multi-model comparison, SMOTE oversampling, GridSearchCV hyperparameter tuning, and a live Streamlit deployment.

---

🚀 Live Demo

🔗 Try the live app here

Enter customer details → Get instant churn prediction + probability score

---

📊 Model Results

Before SMOTE

Model	Accuracy	Precision	Recall	F1 Score
Random Forest	83.71%	64.68%	42.23%	51.10%
XGBoost	82.56%	59.42%	42.49%	49.55%
Decision Tree	77.34%	44.12%	46.63%	45.34%
Logistic Regression	82.40%	65.03%	27.46%	38.62%

After SMOTE

Model	Accuracy	Precision	Recall	F1 Score
Random Forest	79.22%	48.66%	56.48%	52.28%
XGBoost	80.10%	50.64%	51.30%	50.94%
Logistic Regression	70.50%	37.16%	67.10%	47.83%
Decision Tree	74.20%	37.89%	43.78%	40.62%

After GridSearchCV — Tuned Random Forest (Best Model)

Class	Precision	Recall	F1 Score
Retained	0.89	0.85	0.87
Churned	0.49	0.57	0.53

🏆 Best Cross-Validation F1: 86.61% ⚙️ Best Parameters: max_depth: 20, min_samples_leaf: 1, min_samples_split: 2, n_estimators: 200

---

🎯 Objective

Predict whether a bank customer will churn (leave the bank) based on demographic and account features — enabling proactive retention strategies that save revenue.

---

📁 Dataset

• Size: ~7,000 records
• Target: Churn (1 = churned, 0 = retained)
• Features: Age, Balance, Credit Score, Geography, Gender, Tenure, Number of Products, Active Member, Estimated Salary

---

⚙️ Project Workflow

Data Loading & Exploration
        ↓
Exploratory Data Analysis (EDA)
(distributions, correlations, class imbalance check)
        ↓
Statistical Feature Testing
(Chi-Square for categorical, ANOVA for numerical)
        ↓
Data Preprocessing
(Label encoding, MinMaxScaler, IQR outlier removal)
        ↓
Feature Selection
(Sequential Feature Selector — top 9 features)
        ↓
Model Training & Comparison
(Logistic Regression, Decision Tree, Random Forest, XGBoost)
        ↓
SMOTE Oversampling
(Fix class imbalance — boosted Recall from 42% → 56%)
        ↓
GridSearchCV Hyperparameter Tuning
(Tuned Random Forest — Best CV F1: 86.61%)
        ↓
Streamlit Deployment
(Live app with real-time churn prediction)

---

📈 Visualizations

Model Comparison Chart

Confusion Matrix — Tuned Random Forest

Top 10 Feature Importances

---

🔑 Key Findings

• SMOTE significantly improved Recall — Random Forest Recall jumped from 42.23% → 56.48%
• Logistic Regression saw the biggest Recall boost: 27.46% → 67.10% after SMOTE
• Tuned Random Forest achieved best cross-validation F1 of 86.61% via GridSearchCV
• Without SMOTE all models were heavily biased toward predicting "Retained"
• Top churn predictors: Age, Estimated Salary, Active Member status, Country (Germany)

---

🛠️ Tech Stack

Layer	Technology
Language	Python 3.10+
Data Processing	Pandas, NumPy
Visualization	Matplotlib, Seaborn
ML Models	Scikit-learn, XGBoost
Oversampling	imbalanced-learn (SMOTE)
Hyperparameter	GridSearchCV (Scikit-learn)
Feature Selection	Sequential Feature Selector
Statistical Tests	Chi-Square, ANOVA (SciPy)
Deployment	Streamlit Cloud
Environment	Google Colab

---

📁 Project Structure

Customer-Churn-Prediction-using-Machine-Learning/
├── customer_churn_prediction.ipynb   # Full notebook: EDA + modelling + SMOTE + GridSearchCV
├── churn_app.py                      # Streamlit deployment app
├── model.json                        # Trained XGBoost model (native format, 452KB)
├── requirements.txt                  # Python dependencies
├── model_comparision.png             # Bar chart comparing all 4 models
├── confusion_matrix.png              # Confusion matrix for Tuned Random Forest
├── feature_importance.png            # Top 10 feature importances
└── README.md

---

🚀 Run Locally

1. Clone the repo

git clone https://github.com/varun0852/Customer-Churn-Prediction-using-Machine-Learning.git
cd Customer-Churn-Prediction-using-Machine-Learning

2. Install dependencies

pip install -r requirements.txt

3. Run the Streamlit app

streamlit run churn_app.py

4. Or run the notebook

jupyter notebook customer_churn_prediction.ipynb

---

🔮 Next Steps

• Try ensemble stacking of RF + XGBoost
• Experiment with threshold tuning to further improve Churned class F1
• Add SHAP explainability to the Streamlit app

---

👤 Author

Varun — AI/ML Engineer