Precision

Precision answers one blunt question about a classifier: when it says 'yes,' how often is it right? Formally, precision is true positives divided by the sum of true positives and false positives — the fraction of the items flagged positive that truly are positive. If a spam filter marks 100 emails as spam and 90 really are, its precision is 0.9. Precision cares only about the things the model claimed were positive; it ignores the positives it missed. That makes it the metric to watch whenever a false positive is expensive or annoying — flagging a good customer as fraud, blocking a legitimate email, showing an irrelevant ad. High precision means few false alarms: the model is cautious and trustworthy when it commits to a positive call.

Precision is one half of a pair, and it never tells the whole story alone. A model can score perfect precision by being extremely timid — flagging only the one case it is absolutely sure of and staying silent on everything else. That single correct flag gives precision 1.0 while the model misses hundreds of real positives. So precision is always read next to recall, which measures how many of the actual positives were caught. The two trade off: push a model to flag more and recall rises but precision usually falls, because looser flagging drags in false positives. Choosing where to sit on that trade-off is a business decision, not a math one, and it depends entirely on what a false positive versus a missed positive costs you.

The cousin people confuse with precision is recall, and the difference is exactly which mistakes you count. Precision divides true positives by everything the model flagged positive, so its enemy is the false positive — the false alarm. Recall divides true positives by everything that was actually positive, so its enemy is the false negative — the miss. Precision asks, 'of what I flagged, how much was right?' Recall asks, 'of what mattered, how much did I catch?' A cancer screen wants high recall, since missing a real case is deadly, even at the cost of false alarms; a spam filter wants high precision, since junking a real email is worse than letting one spam through. The F1 score blends the two into a single number when you need one figure that respects both.

Accuracy is a third relative, and it is the one that quietly lies on imbalanced data. Accuracy is simply the share of all predictions that were correct, positive and negative alike. That sounds ideal until the classes are lopsided. If one in a thousand transactions is fraud, a model that predicts 'not fraud' every time scores 99.9% accuracy while catching zero fraud — useless. Precision and recall expose that, because they focus on the rare positive class the model is supposed to find. This is why fraud, disease, and churn problems are judged on precision and recall rather than headline accuracy. When someone quotes only accuracy on a rare-event problem, treat it as a red flag and ask for the precision and recall behind it.

Use precision when the cost of a false positive is what keeps you up at night. Set the decision threshold deliberately: raising it makes the model flag only its most confident cases, lifting precision while lowering recall; lowering it does the reverse. Report precision alongside recall — never one alone — and, when you need a single number, use the F1 score or a precision-at-fixed-recall target. Watch precision on the class you care about, not a blurred average, and re-check it as data drifts, because a filter that was precise last quarter can decay as spam or fraud tactics change. In marketing, precision governs things like lookalike targeting and lead scoring: a high-precision 'likely buyer' list wastes less budget, even if it reaches fewer people.

The failures cluster around reading precision in isolation. A model tuned for precision alone can look brilliant while quietly missing most of the real positives, because timid flagging inflates the score. Quoting accuracy instead of precision on imbalanced data hides a model that never finds the rare class. Ignoring the threshold — accepting the default 0.5 — leaves precision to chance when the business needs it dialed in. And forgetting to monitor drift lets a once-precise model rot. The discipline is to name the cost of a false positive, tune the threshold to hit the precision you need, read it beside recall, and keep watching it over time. Precision is a promise about the model's 'yes,' and promises have to be maintained.

In statistics and machine learning, 'precision' — also called positive predictive value — measures the correctness of positive predictions, standing opposite 'recall' in information-retrieval evaluation.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where precision is a core concern: