This project follows the methodology of the article “Should This Loan be Approved or Denied?”, identifying indicators of potential risk in SBA loan data, building and validating a logistic regression model, and applying the model to make lending decisions.

The datasets are not included in the repository.

You can download them from here. Take the entire Data folder and put it inside your project folder.

• States like Florida, D.C., and Georgia have the highest loan default rates (>23%), indicating higher credit risk.
• Mountain and Northern Plains states like Montana, Wyoming, and Vermont have the lowest default rates (<8%), showing relatively safer lending environments.

• Real Estate, Finance, and Transportation show the highest default rates (~28–29%), indicating high-risk sectors.
• Mining, Health Care, and Management have the lowest default rates (<11%), suggesting more stable lending.

CHGOFF and PIF Quartiles

• Smaller loans are more likely to default.
• Large loans are often safer: 75% quartile for repaid 255K > defaulted $140K.
• DisbursementGross is a strong predictor: Clear difference between distributions for defaulted vs repaid loans.

• New businesses default slightly more than existing businesses.
• NewExist is not a strong predictor.

• Real-estate loans are far safer, with 12-13x lower default risk.
• RealEstate is a strong predictor.

• Loans active during the Great Recession have a higher default rate (32.3%) than loans that were not active during the Recession (16.9%).
• Recession shows some economic impact and may support other predictors in the model.

• Defaulted loans tend to have a slightly lower median SBA portion than fully paid loans.
• Both distributions are similar, suggesting SBA coverage has limited direct effect on default rates.
• Portion is not a very strong predictor, though it may support other predictors in the model.

Type 3 Test Results

• RealEstate and Portion are highly significant, meaning they strongly affect default probability.
• Recession is mildly significant.
• New and DisbursementGross are not significant, so they may not contribute meaningfully once other variables are included.

Logit Regression Results

Changes in results:

• Model fit: Pseudo R² remained nearly unchanged (0.174 → 0.173), indicating no meaningful loss in explanatory power.
• Log-Likelihood: Slight change from -540.85 to -541.38 despite removing 2 predictors, confirming they contributed little.
• Significance: The remaining predictors retained their effect sizes and statistical significance.

Dropping non-significant variables simplified the model without reducing predictive strength, improving interpretability.

• There is an optimal cutoff value around 0.5 cutoff where the misclassification rate is minimized.

Confusion Matrix

Misclassification rate at cutoff 0.5: 32.16%

• High Risk → 31 correctly predicted defaults, 14 incorrect fully paid predictions.
• Low Risk → 682 correctly predicted fully paid, 324 missed defaults.

This confirms that the 0.5 cutoff is close to optimal based on the misclassification rate curve.

ROC Curve:

Summary:

The model has a fair ability to classify positives and negatives, with an AUC of 0.735 indicating moderate predictive performance.

Loan Applications

Based on the predictions from the logistic regression model, the classification was accurate and interpretable. The model successfully identified low and high-risk loans using meaningful features.