A fraud model reports 99% accuracy when only 1% of transactions are fraudulent. How would you evaluate it?

A classifier that always predicts legitimate already achieves 99% accuracy, so inspect the confusion matrix, fraud-class precision and recall, and precision-recall behavior across thresholds. Estimate the costs of missed fraud and unnecessary investigations, then select a threshold that fits operational capacity. Use a time-based evaluation if production predicts future transactions. Check calibration when scores guide decisions, and examine important customer segments. Report results at the chosen threshold rather than presenting one impressive aggregate metric without its business consequences.