Personal loan applications

What it is good for

• Tabulate default rates by credit-score band and purpose.
• Fit and interpret a logistic regression (odds ratios) for default.
• Evaluate a classifier with a confusion matrix, sensitivity/specificity, ROC/AUC at different thresholds.
• Discuss why only approved loans appear (selection) and what that means for conclusions.

Methods: contingency tables, logistic regression, odds ratios, ROC curve, AUC, confusion matrix, calibration.

Data quality notes

• Overall default rate is about 13% — an imbalanced outcome.
• employment_years is missing for 61 self-employed applicants. In this simulation they were chosen at random, so the missingness is unrelated to the other variables (MCAR) — real data are rarely that kind.
• Credit score, DTI and interest rate are correlated (the rate is priced from the score).
• Only approved applications are included, so the data cannot show how rejected applicants would have performed.

Provenance and licence

Monthly income is log-normal (median ≈13 500 SAR). Credit score is normal (mean ≈690, sd 65) shifted slightly by income, clamped to 300–850. Debt-to-income falls with credit score. Loan amount ≈ income × gamma-distributed multiple, rounded to 1 000 SAR; term is more often 60 months for larger loans; the interest rate is priced from the credit score. Default (0/1) is Bernoulli with a logistic model in credit score, DTI, 60-month term, employment years, business purpose and loan-to-income (coefficients exported as LOANS_TRUE_MODEL). Employment years are then blanked for 61 self-employed applicants.

Synthetic data · CC0-1.0 (synthetic data generated by Statistics Quest) · generated deterministically (seed 20241010). It does not describe real people or organisations; any resemblance is coincidental.