PRODUCTION SCENARIO
A payments company trains a binary fraud classifier. Fraud is 0.8 percent of transactions, and a model that predicts not-fraud for every transaction reports 99.2 percent accuracy. The team must compare candidate models on how well they rank the rare fraud cases across every possible threshold.
Which evaluation metric is MOST appropriate?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Area under the precision-recall curve, computed across all thresholds
AUPRC summarises precision against recall across every threshold and does not count true negatives, so it stays informative when 99.2 percent of rows are negative. AUC-ROC is also threshold-free, but its false positive rate is diluted by that same negative majority. F1 at 0.5 and balanced accuracy each score one fixed threshold, so neither compares how well a model ranks the rare class.