A fraud team is choosing an evaluation metric. Fraud is 0.2% of transactions. Two models have identical ROC-AUC of 0.97, but Model P has a PR-AUC of 0.61 and Model Q a PR-AUC of 0.34.
Which metric should decide, and which model wins?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
ROC plots recall against false-positive rate, and with 99.8% legitimate transactions the denominator is enormous — thousands of false alarms barely move the curve. Precision–recall asks the question the fraud team cares about: of the transactions we block, how many were fraud? Model P delivers far more precision at every recall level, so it wins decisively despite the identical ROC-AUC.
Both metrics are threshold-free. ROC-AUC is simply insensitive to the class the business cares about when the positive class is rare.
A model that never flags anything scores 99.8% accuracy. It tells you nothing about fraud detection.
PR-AUC is higher when precision is higher — i.e., fewer false alarms per catch. Model Q raises more false alarms, not fewer.