Customer Churn Prediction — Lloyds Banking Group

A complete end-to-end machine learning project predicting customer churn using demographic, transactional, and behavioural data.

---

Project Overview

Item	Detail
Problem	Binary classification — predict if a customer will churn
Dataset	1,000 customers across 5 data sources
Best Model	Logistic Regression (Tuned)
Best Recall	46.3%
Best F1 Score	27.9%
Best AUC-ROC	52.9% (Random Forest)

---

Project Structure

churn_prediction_project/
├── data/
│   └── raw/                    ← original data (not uploaded)
├── notebooks/
│   ├── 01_EDA_Churn.ipynb      ← exploratory data analysis
│   └── 02_Preprocessing_Modelling.ipynb ← ML pipeline
├── reports/
│   ├── 01_churn_distribution.png
│   ├── 02_age_analysis.png
│   ├── 03_categorical_churn.png
│   ├── 04_transaction_analysis.png
│   ├── 05_online_activity_churn.png
│   ├── 06_service_churn.png
│   ├── 07_correlation_heatmap.png
│   ├── 08_model_comparison.png
│   └── 09_feature_importance.png
├── .gitignore
├── README.md
└── requirements.txt

---

Key Findings

EDA Findings

• Dataset is 80/20 imbalanced — 796 stayed, 204 churned
• LoginFrequency is the strongest predictor (correlation: -0.08)
• TotalSpent and NumTransactions have 0.90 correlation — multicollinearity detected
• 332 customers never contacted customer service (structural zeros)
• Categorical features (Gender, MaritalStatus) show weak churn signal

Model Results

Model	Accuracy	Precision	Recall	F1 Score	AUC-ROC
Logistic Regression	51.0%	20.0%	46.3%	27.9%	47.4%
Random Forest	63.0%	20.0%	26.8%	22.9%	52.9%
XGBoost	55.5%	14.7%	24.4%	18.3%	47.6%

Why Logistic Regression was selected

• Highest Recall (46.3%) — catches most actual churners
• No overfitting — Train-Test gap < 10%
• Interpretable — required by FCA banking regulations
• Random Forest and XGBoost showed severe overfitting before tuning

---

Top Churn Risk Factors

Rank	Feature	Direction	Business Meaning
1	LoginFrequency	↓ decreases churn	More logins = loyal customer
2	AvgSpent	↑ increases churn	Higher spenders still churn
3	DaysSinceLogin	↑ increases churn	Inactivity = churn signal
4	IncomeLevel_Low	↑ increases churn	Price-sensitive segment
5	MainInteractionType_Complaint	↑ increases churn	Complaints predict leaving

---

How to Run This Project

1. Clone the repository

git clone https://github.com/YOUR_USERNAME/churn_prediction_project.git
cd churn_prediction_project

2. Create virtual environment

python -m venv venv
venv\Scripts\activate        # Windows
source venv/bin/activate     # Mac/Linux

3. Install dependencies

pip install -r requirements.txt

4. Add your data

Place Customer_Churn_Data_Large.xlsx in data/raw/

5. Run notebooks in order

01_EDA_Churn.ipynb              → EDA and findings
02_Preprocessing_Modelling.ipynb → ML pipeline

---

Technologies Used

Tool	Purpose
Python 3.13	Core language
Pandas 3.0.1	Data manipulation
NumPy 2.4.3	Numerical computing
Matplotlib 3.10	Visualisation
Seaborn 0.13.2	Statistical plots
Scikit-learn 1.8.0	ML models and preprocessing
XGBoost 3.2.0	Gradient boosting
Jupyter Notebook	Development environment

---

Business Recommendations

1. Deploy Logistic Regression for monthly customer scoring
2. Three-tier risk segmentation — High/Medium/Low risk
3. Estimated impact — £14M+ annual revenue protected at scale
4. Key action — Re-engage customers with <15 logins/month
5. Data improvement — Add account tenure and product holdings

---

Author

Samitha Sandaruwan
Aspiring Data Scientist
GitHub

---

Acknowledgements

Project developed as part of Lloyds Banking Group Data Science Job Simulation.