Credit-Card-Behavior-Score-Prediction

Credit Card Default Prediction

📌 Project Overview

Credit cards provide convenient short-term credit, but also carry the risk of customer default. This project develops a binary classification model to predict whether a customer will default in the next billing cycle based on historical behavioral and financial data.

By identifying high-risk customers early, financial institutions can reduce potential losses, improve decision-making, and strengthen credit risk management.

Target variable:

• next_month_default = 1 → Customer defaults next month
• next_month_default = 0 → No default

---

🎯 Objective

• Build a robust predictive model for credit card default.
• Prioritize identifying defaulters (high recall) over avoiding false positives.
• Ensure financial interpretability by engineering meaningful behavioral features.

---

📂 Repository Structure

project-name/
├── data/                # Train/test datasets (or sample data if large)
├── notebooks/           # Jupyter notebook(s) with full code
├── reports/             # Final project report and figures
├── requirements.txt     # Python dependencies
├── README.md            # Documentation

---

🔎 Data and Preprocessing

• Dataset: 25,247 rows × 27 columns (demographics, credit limit, bill amounts, payment history).

• Steps performed:

  • Dropped Customer_ID and duplicates.
  • Imputed missing age values with median.
  • Fixed invalid categorical values in education and marriage.
  • Log-transformed skewed features like LIMIT_BAL.

---

📊 Exploratory Data Analysis (EDA)

Key findings:

• Defaulters show higher payment delays (PAY_0 to PAY_4).
• Lower credit limits were associated with higher defaults.
• Financial behavior indicators such as repayment ratios and credit utilization strongly influenced default risk.

---

🛠️ Feature Engineering

Engineered features to capture financial behavior:

• credit_utilization → Bill amount ÷ Credit limit
• delay_months → Count of delayed months
• repayment_ratio → Total payments ÷ Total bills
• avg_delay → Average payment delay
• delinquency_streak → Longest streak of delayed months

---

🤖 Model Building

Models trained:

• Logistic Regression
• Random Forest
• LightGBM
• XGBoost (best performer)

Class imbalance handled using SMOTE. Threshold tuning (0.11) emphasized recall and F2 score.

---

📈 Model Performance

Model	Precision (1)	Recall (1)	F1 Score	F2 Score	AUC
Logistic Regression	0.40	0.64	0.49	0.57	–
Random Forest	0.27	0.84	0.41	0.59	0.75
LightGBM	0.33	0.76	0.46	0.61	0.77
XGBoost (best)	0.28	0.86	0.42	0.61	0.76

👉 XGBoost was selected as the final model due to its high F2 score (0.607) and recall, aligning with the business goal of minimizing losses from missed defaulters.

---

🚀 How to Run

1. Clone the repository:

   git clone https://github.com/username/project-name.git
   cd project-name

2. Install dependencies:

   pip install -r requirements.txt

3. Open the Jupyter notebook:

   jupyter notebook notebooks/project_notebook.ipynb

---

📜 Conclusion

• Developed a predictive model for credit card default risk.
• Created financially meaningful features improving interpretability.
• Addressed class imbalance using SMOTE.
• Tuned classification thresholds to prioritize recall.
• XGBoost emerged as the best model, helping financial institutions proactively manage risk.

---

🛠️ Technologies Used

• Python (Pandas, NumPy, Matplotlib, Seaborn)
• Scikit-learn
• XGBoost, LightGBM, Random Forest
• Jupyter Notebook
• SMOTE (Imbalanced-learn)

---

📑 Report

The full detailed analysis, methodology, and findings are available in reports/final_report.pdf.

---