Customer Churn Prediction 📊

An end-to-end machine learning system that predicts whether a bank customer will churn —
built with Neural Networks and deployed as an interactive Streamlit web application.

---

📌 Project Overview

Customer churn costs banks millions annually. Identifying at-risk customers before they leave allows businesses to take targeted retention actions — saving revenue and improving customer relationships 🎯

This project answers the question:

"Given a customer's demographic and banking profile, how likely are they to leave?"

---

📚 Table of Contents

• 📌 Project Overview
• 🚀 Live Demo
• 📂 Project Structure
• 🛠️ Tech Stack
• 📊 Dataset Description
• 📈 Key Insights from EDA
• ⚙️ Data Preprocessing
• 🤖 Model Training & Performance
• 🧠 Neural Network Architecture
• 🔮 Inference Pipeline
• 🚀 Application Features
• ▶️ Run Locally
• 💡 Business Recommendations
• 🔭 Future Improvements
• 👩‍💻 Author

---

🚀 Live Demo

⚠️ App may be sleeping — Streamlit free tier may hibernate. Click Live App and wait ~30s.

---

📂 Project Structure

Customer-Churn-Prediction/
│
├── app/
│   └── app.py                        # Streamlit web application
│
├── data/
│   ├── raw/
│   │   └── Churn_Modelling.csv       # Original dataset
│   └── processed/
│       └── Churn_processed.csv       # Cleaned and encoded dataset
│
├── models/
│   ├── churn_prediction_model.h5     # Trained Neural Network
│   ├── scaler.pkl                    # Fitted StandardScaler
│   └── training_columns.pkl          # Column list for consistent inference
│
├── notebooks/
│   └── churn_analysis.ipynb          # Full analysis notebook
│
├── src/
│   ├── preprocessing.py              # Data loading and preprocessing functions
│   ├── train_model.py                # Model building and training functions
│   └── evaluate.py                   # Model evaluation functions
│
├── requirements.txt
└── README.md

---

🛠️ Tech Stack

🏷️ Category	🔧 Tools
🐍 Language	Python 3.11
🧠 Deep Learning	TensorFlow, Keras
📊 ML & Data	Scikit-learn, Pandas, NumPy
📈 Visualization	Matplotlib, Seaborn
🚀 Deployment	Streamlit
💾 Model Serialisation	Joblib

---

📊 Dataset Description

The dataset used is the Credit Card Customer Churn dataset from Kaggle.

• Total Records: 10,000 customers
• Features: 14 columns (RowNumber, CustomerId, Surname dropped before modelling)
• Target Variable: Exited (1 = Churned, 0 = Stayed)
• Class Distribution: 79.6% Stayed / 20.4% Churned — handled via class weights

🔑 Feature	📝 Description
CreditScore	💳 Customer credit score
Geography	🌍 Customer location (France, Germany, Spain) — one-hot encoded
Gender	👤 Male / Female — one-hot encoded
Age	🎂 Customer age
Tenure	📅 Years with bank
Balance	💰 Account balance
NumOfProducts	📦 Number of bank products
HasCrCard	💳 Has credit card (1 = Yes, 0 = No)
IsActiveMember	✅ Active member status (1 = Yes, 0 = No)
EstimatedSalary	💵 Customer estimated salary

---

📈 Key Insights from EDA

🔍 Insight	💼 Business Implication
📉 20.4% overall churn rate	Significant revenue at risk — retention efforts needed
🌍 Germany has highest churn	Target German customers with special retention offers
👴 Age 50–60 has 56.2% churn rate	Highest risk group — prioritise retention for middle-aged customers
🧑 Age 40–50 has 34.0% churn rate	Second highest risk group — early intervention recommended
🧒 Age <30 has only 7.5% churn rate	Most loyal segment — leverage for referrals and upselling

Fig 1 — Churn Distribution: 79.6% stayed vs 20.4% churned	Fig 2 — Germany churns at 32.4%, nearly double non-German customers
Fig 3 — Age 50–60 has the highest churn rate at 56.2%; under-30s are the most loyal at 7.5%

---

⚙️ Data Preprocessing

🔢 Step	⚙️ Detail
🗑️ Drop columns	RowNumber, CustomerId, Surname removed
🔤 One-Hot Encoding	Geography and Gender encoded using pd.get_dummies(drop_first=True)
💾 Column list saved	training_columns.pkl saved to ensure consistent inference
✂️ Train-Test Split	80% train / 20% test, random state = 1
📏 Feature Scaling	StandardScaler — fit_transform on train, transform on test

---

🤖 Model Training & Performance

Multiple models were evaluated with class_weight='balanced' applied to handle the 80/20 class imbalance.

🤖 Model	🎯 Accuracy
📈 Logistic Regression	~72%
🌳 Decision Tree	~76%
✅ Neural Network	~80%

The Neural Network was selected as the final model as it significantly outperforms the baseline models and captures complex, non-linear patterns in customer behaviour that simpler models miss. Decision Trees tend to overfit, and Logistic Regression assumes linear relationships — both limitations the Neural Network overcomes with its layered architecture and Dropout regularisation.

Fig 4 — (Top) Training vs Validation Loss & Accuracy over 50 epochs  |  (Bottom) Confusion Matrix & Permutation Feature Importance

The model converges steadily across 50 epochs with no signs of overfitting. The confusion matrix confirms strong performance on the majority class, while the permutation feature importance chart reveals that NumOfProducts and Age are the most influential predictors of churn.

---

🧠 Neural Network Architecture

🏗️ Layer	⚙️ Configuration
➡️ Input Layer	11 features
🔵 Dense Layer	64 neurons (ReLU)
🔄 Dropout	0.3
🔵 Dense Layer	32 neurons (ReLU)
🔄 Dropout	0.3
🎯 Output Layer	1 neuron (Sigmoid)

⚡ Training Config	📊 Value
📉 Loss Function	Binary Crossentropy
⚙️ Optimizer	Adam
🔁 Epochs	50
📦 Batch Size	32
⚖️ Class Weights	Applied to address 80/20 imbalance

---

🔮 Inference Pipeline

At inference time three saved assets are loaded:

1. churn_prediction_model.h5 — the trained neural network
2. scaler.pkl — the fitted StandardScaler
3. training_columns.pkl — the exact column list after one-hot encoding

This guarantees new customer data is encoded identically to how the training data was processed — no silent column mismatches.

---

🚀 Application Features

The deployed Streamlit dashboard includes:

• 🎛️ Customer Input Panel — credit score, age, tenure, balance, number of products, credit card status, active membership, estimated salary, geography, and gender
• 🤖 Real Model Inference — predictions from the trained neural network
• 🚦 Three Risk Tiers — High (>50%), Medium (30–50%), Low (<30%) with visual progress bar
• 💡 Retention Suggestions — contextual tips based on the customer's specific risk factors

---

▶️ Run Locally

# 1. Clone the repository
git clone https://github.com/mysticalayushi/Customer-Churn-Prediction.git
cd Customer-Churn-Prediction

# 2. Install dependencies
pip install -r requirements.txt

# 3. Launch the Streamlit app
streamlit run app/app.py

Or explore the analysis notebook:

jupyter notebook notebooks/churn_analysis.ipynb

---

💡 Business Recommendations

1. 🎯 Target German customers with special retention offers as they show higher churn probability
2. 💤 Engage inactive members through loyalty programs, personalised emails, or incentives
3. 👴 Focus on customers aged 50–60 as this group has the highest churn rate at 56.2%
4. 🧑 Monitor customers aged 40–50 as they represent the second highest risk group at 34.0%
5. 🎁 Introduce personalised financial products based on customer behaviour and risk score

---

🔭 Future Improvements

• 🌲 Improve model performance using Gradient Boosting or XGBoost
• 🔧 Perform hyperparameter tuning to further optimise prediction accuracy
• 📊 Incorporate additional customer behaviour features such as transaction frequency
• 🧪 Implement A/B testing strategies to evaluate churn reduction campaigns
• ⚡ Build a real-time data pipeline for automated churn monitoring
• 🔍 Add SHAP values for deeper model explainability

---

📋 Project Information

📌 Field	📝 Detail
👩‍💻 Created by	Ayushi Rai
🧠 Model	Neural Network (TensorFlow / Keras)
🎯 Test Accuracy	~80%
📊 Dataset	Credit Card Customer Churn — Kaggle (10,000 records)
📅 Date	March 2026

---

👩‍💻 Author

Ayushi Rai

---

_{If you found this project helpful, consider giving it a ⭐ on GitHub!}

---

⬆️ Back to Top