Logistic regression and gradient boosting models trained on restaurant-level processing volume and transaction-hours features. Evaluation on a 25% test split, with calibration assessed against actual default rates.