What is precision?
Precision is the share of a model’s positive predictions that turn out to be correct. If a classifier flags 100 invoices as duplicates and 80 of them really are duplicates, its precision is 0.80. The formula is true positives divided by the sum of true positives and false positives. In medicine and diagnostics the same quantity is called positive predictive value (PPV), the probability that a patient with a positive test result actually has the condition.
Precision describes how trustworthy a positive prediction is. It says nothing about the positives the model missed, which is what recall measures. The two pull against each other: a stricter confidence threshold raises precision and lowers recall, and a looser one does the opposite. The F1 score combines them when a single number is needed.
Why precision matters
Every false positive has a cost, and precision is the metric that tracks it. A low-precision model buries the people who act on its output in alerts that lead nowhere, and they soon learn to ignore it. Precision matters most when each flagged case triggers expensive human work or an action that annoys a customer, such as a blocked payment or a hold on a shipment.
Precision also depends on prevalence, the share of positive cases in the data. A test that is right 99 times out of 100 on both positives and negatives still produces mostly false alarms when the condition affects one case in a thousand. Accuracy figures hide this problem, so ask for precision measured on data with a realistic class balance.
Precision in practice
A building products distributor runs a model that flags probable duplicate supplier invoices before payment. The accounts payable team can investigate about 30 flagged invoices a day. At the first threshold the model flagged 120 a day and only 25 were genuine duplicates, a precision near 0.20. Raising the threshold cut the daily flags to 35, of which 28 were genuine, a precision of 0.80. The team now clears the queue every day.
What to watch for
- Ask what confidence threshold a quoted precision figure was measured at, because a different threshold gives a different number.
- Check that the test data had the same prevalence as production, since precision falls as the positive class becomes rarer.
- Track precision per segment, such as by supplier or region, because an acceptable average can hide a segment where the model is almost always wrong.
- Agree with the business owner how many false positives the review team can absorb each day, then set the threshold to match.