Accuracy is usually the wrong metric
On an imbalanced problem — 1% fraud — a model predicting 'never fraud' scores 99% accuracy and is useless. Precision (of the flagged items, how many were right) and recall (of the real positives, how many we caught) expose the trade-off; F1 blends them.
Threshold-free views
ROC-AUC summarises ranking quality across every threshold, while precision–recall AUC is more informative when positives are rare. Choose the operating threshold from business cost, not the default 0.5.
Calibration and slices
A calibrated model that says 0.7 is right about 70% of the time; check with a reliability diagram. Always evaluate per-slice — by geography, device or demographic — because a strong average can hide a failing subgroup.