precision
/ pruh-SIZH-un /
Precision answers a pointed question: when the model says yes, how often is it actually right? Picture a spam filter. Of all the messages it threw into the spam folder, what share were truly spam? If it dumped 50 emails there and 45 were genuine junk, its precision is 45/50 = 90 percent. The five real emails it wrongly trashed are the cost of getting it wrong.
Formally, precision is the number of true positives divided by everything the model called positive — true positives plus false positives. It cares only about the predictions where the model said yes, and ignores the cases it left alone. High precision means few false alarms: when it raises its hand, you can trust it.
Precision matters most when a false alarm is expensive or annoying. You want a spam filter that almost never buries a real message; you want a cancer-confirmation test that rarely scares a healthy person. But precision is only half the story — a model can score perfect precision by being absurdly cautious, flagging just one obvious case and staying silent on everything else. To see what it missed, you must pair precision with recall.
A model flags 8 transactions as fraud. Investigating them, 6 turn out to be real fraud and 2 are legitimate purchases. Precision = 6/8 = 75%. Three-quarters of its alarms were worth chasing; one-quarter wasted an investigator's time.
Precision = of the alarms raised, the fraction that were genuine.
Precision and recall trade off against each other. Crank the model to flag fewer, surer cases and precision rises but recall falls; flag more eagerly and recall rises but precision falls. Reporting one without the other is a classic way to make a model look better than it is.