Credit Default Prediction

IMT 572: Introduction to Data Science – Final Project

Overview

This project applies data science techniques to the UCI Credit Card Default Dataset, exploring predictors of credit default and evaluating classification models. The workflow included:

• Exploratory Data Analysis (EDA) in R and Tableau
• Regression-based exploratory modeling (Logit & Probit)
• Classifier training & evaluation (kNN, SVM, Random Forest)
• Visualization & dashboarding with Tableau

The goal was to understand key drivers of default, reduce dimensionality, and identify the most accurate predictive model.

---

Data

• Source: Credit dataset (CSV)
• Size: 30,000 records
• Key variables:
  • LIMIT_BAL – credit limit
  • SEX, EDUCATION, MARRIAGE, AGE – demographic variables
  • BILL_AMT1–6 – historical bill amounts
  • PAY_AMT1–6 – repayment amounts
  • PAY_0–PAY_6 – repayment status (history of delays)
  • default.payment.next.month – target (binary: default or not)

---

Exploratory Analysis

Correlation Insights

• BILL_AMT1–6 are highly correlated → risk of multicollinearity. Only BILL_AMT1 retained.
• Payment amounts (PAY_AMT1, PAY_AMT3) provide unique predictive power.
• LIMIT_BAL shows low correlation with others but is a key protective factor against default

Demographic Insights (Tableau)

• Females show higher credit usage and default rates than males, especially ages 25–30.
• Singles exhibit slightly higher defaults compared to married individuals.
• Younger individuals (late 20s/early 30s) have higher bills, payments, and defaults.
• Higher education (university-level) correlates with higher bills and greater default risk

---

Regression Analysis

Logit & Probit Models

• Significant predictors: LIMIT_BAL, PAY_AMT1, PAY_AMT3, BILL_AMT1, PAY_0, PAY_2, PAY_3.
• LIMIT_BAL → negative effect (higher limits = lower default probability).
• PAY_AMT1, PAY_AMT3 → negative effect (larger payments reduce default).
• PAY_0, PAY_2, PAY_3 → positive effect (late payments increase default).
• Probit slightly outperforms Logit (lower AIC)

---

Classifier Models

Model	Accuracy
Random Forest	81.99%
SVM (Linear)	81.74%
SVM (Radial)	81.65%
kNN (k=19)	79.63%

• Random Forest is the best-performing model.
• SVM Linear offers comparable performance with potential computational efficiency benefits.
• kNN performs reasonably but lags behind ensemble and kernel methods

---

Visualization

Interactive dashboards were built in Tableau (Final project.twb), highlighting:

• Default rates by demographics
• Credit usage patterns
• Payment delays and amounts
• Predictive variable comparisons

---

Key Takeaways

• Payment history (PAY_0, PAY_2, PAY_3) is the strongest predictor of default.
• Credit limit (LIMIT_BAL) is protective against default risk.
• Random Forest achieves the highest classification accuracy.
• SVM (Linear) is a strong alternative when runtime efficiency matters.

---

Files in Repository

• credit_dataset.csv – Dataset used for analysis
• Final Project.pdf – Detailed report with results and interpretations
• Final project.twb – Tableau workbook with visualizations

---