Why is accuracy misleading for fraud detection, and what metrics instead?

WHAT THIS TESTS: This question probes whether you understand the accuracy paradox and class imbalance. Interviewers want to see that you know why a high accuracy score can mask a useless model when the positive class is rare, and that you can select metrics aligned with business impact rather than mathematical convenience.

A GOOD ANSWER COVERS: First, the numerical trap. With only one percent fraud prevalence, a trivial classifier that labels every transaction as legitimate achieves ninety-nine percent accuracy while catching zero actual fraud. Second, the remedy. You should propose precision to measure how many flagged transactions are truly fraudulent, recall to measure how much fraud you catch, and the F1 score to balance them. Third, you should mention AUC-PR because precision-recall curves are more informative than ROC curves under severe imbalance. Fourth, tie metrics to business costs. High recall minimizes missed fraud, while high precision reduces costly false alarms and operational alert fatigue.

COMMON WRONG ANSWERS: A major red flag is claiming accuracy is still fine because it is easy to explain. Another is suggesting ROC-AUC as the sole fix without acknowledging that ROC can look artificially optimistic with extreme imbalance. Proposing only precision without recall is also weak because a model that never flags anything has undefined or misleading precision. Finally, ignoring the cost asymmetry between false negatives and false positives signals a lack of product thinking.

LIKELY FOLLOW-UPS: Expect the interviewer to ask how you would handle the imbalance algorithmically, such as via class weighting, oversampling with SMOTE, or threshold tuning. They may also ask how you would set an operating point on a precision-recall curve given a fixed investigation budget, or how you would explain to executives why a ninety-nine percent accurate model is worthless.

ONE CONCRETE EXAMPLE: Imagine one million transactions with ten thousand fraudulent. A naive model predicts zero fraud and gets ninety-nine percent accuracy. In contrast, a real model flags twelve thousand transactions, correctly identifying eight thousand fraud cases. Its precision is sixty-seven percent and recall is eighty percent. The F1 score is roughly seventy-three percent, and the AUC-PR might be point eight five. These numbers tell a truthful story that ninety-nine percent accuracy completely hides.