Microloan Default

A lender needs to catch high-risk applicants before the money goes out. Both mistakes cost — approving a defaulter loses the principal, rejecting a good borrower loses a customer. The target is default within 90 days, and only 10.9% of applicants default, so the class is heavily imbalanced. That imbalance shaped every decision.

I split train and test by application date, not randomly, so the model is judged on predicting future borrowers from past ones — and I fit every transform on the training window only, so nothing leaks backward. Then the unglamorous work that actually decides quality: missingness flags, sentinel-code detection (the 999 placeholder), logical-consistency checks, and IQR winsorization instead of dropping rows.

Across Logistic Regression, HistGradientBoosting, and Random Forest, Logistic Regression won on the out-of-time test set (ROC-AUC 0.93) and stayed interpretable. I read importance three ways — coefficients, permutation, SHAP — so the drivers agreed rather than resting on one method. Savings balance dominated; unemployed applicants defaulted at 50%, against 6% for the formally employed. I tuned the threshold toward recall, since a missed defaulter usually costs more than a second look at a good one.

A model is only useful if someone can act on it — so I derived a plain screening rule (low savings, high debt-to-income, high loan-to-income) that flags 8.4% of applicants whose actual default rate is 54.6%, against 7.3% for the rest. That’s a 7.5x concentration of risk in a small, explainable segment a credit officer can apply without running a model live.

This is a synthetic teaching dataset, and one feature — savings balance — carries most of the signal, which inflates the AUC beyond what real lending data would give. On real data I’d expect lower scores, and I’d guard against any single feature dominating. The value here is the method, not the number.