Precision and Recall

Precision and recall are two metrics that describe different ways a classification model can fail. Precision measures how often the model is right when it flags something — if it flags 100 transactions as fraudulent and 80 actually are, precision is 80%. Recall measures how much it catches — if there were 200 fraudulent transactions and the model flagged 160 of them, recall is 80%. Improving one typically worsens the other: a model tuned to catch more fraud will also generate more false alarms, and a model tuned to avoid false alarms will miss more fraud. Where to set that threshold is not a technical question.

The precision-recall trade-off is a business decision in disguise. False positives and false negatives have different costs depending on context — blocking a legitimate customer transaction costs differently than missing a fraudulent one, and both carry reputational and financial consequences. Technical teams often set default thresholds without input from the business on what those costs actually are. When executives ask only about accuracy and ignore precision and recall, they're letting the technical team make an implicit business judgment about which type of mistake is more acceptable.