💳 Credit Scoring Model for Default Prediction

CodeAlpha Data Science Internship - Task 1

---

📌 Project Overview

The goal of this project is to build a robust machine learning pipeline to predict the probability of credit card default. Using the Taiwan Credit Card dataset, I developed a model that helps financial institutions identify high-risk applicants, shifting the focus from simple accuracy to high-sensitivity risk detection.

🚀 The Result

By moving from Logistic Regression to a Tuned XGBoost model, I improved the detection of defaulters (Recall) from 0% to 86%, achieving an ROC-AUC score of 0.866.

---

🛠️ Tech Stack & Tools

• Language: Python 3.x
• Libraries: Pandas, NumPy, Scikit-Learn, XGBoost
• Visualization: Matplotlib, Seaborn
• Environment: Google Colab / Jupyter Notebook

---

📂 Project Workflow

1. Data Cleaning & Preprocessing

• Handling Missing Values: Imputed MonthlyIncome with the median and NumberOfDependents with the mode.
• Outlier Correction: Addressed "masking codes" (values 96-98) in payment history columns to ensure data integrity.
• Feature Scaling: Applied StandardScaler to normalize features for linear model performance.

2. Feature Engineering

I engineered new features to capture deeper behavioral insights:

• TotalLate: Aggregated count of all late payments.
• SeverityScore: A weighted penalty system where 90-day delays are weighted 3x more than 30-day delays.
• HighUtilizationFlag: A binary indicator for customers exceeding their credit limit.

3. Model Training & Comparison

I evaluated three different algorithms to determine the best business solution:

Metric	Logistic Regression	Random Forest	Tuned XGBoost
ROC-AUC	0.8255	0.8383	0.8666
Recall (Class 1)	0.00	0.17	0.86
Business Value	Failed to find risk	Missed 83% of risk	Caught 86% of risk

---

🔍 Key Findings (Feature Importance)

The XGBoost model identified that the most significant predictors of default are:

1. TotalLate: Previous payment behavior is the #1 predictor.
2. SeverityScore: The intensity of delays significantly impacts the risk profile.
3. Revolving Utilization: Customers using a high percentage of their credit limit are at higher risk.

---

🏆 Conclusion

In credit risk modeling, Recall is more important than Accuracy. A bank would rather "double-check" a safe customer than miss a customer who will default on a $50,000 loan. This project successfully optimized for Recall (86%), providing a safe and reliable tool for credit assessment.

---

👤 Intern Details

• Name: [Tahir Ahmad]
• Domain: Data Science
• Organization: CodeAlpha
• Task: 1 (Credit Scoring)

---