precision

Precision (Positive Predictive Value)

Of all the items a model predicted as positive, the fraction that were actually positive — the model's ability to avoid false alarms.

When you'd see it: Any classification task where false positives are costly: fraud alerts (investigating non-fraudulent transactions wastes money), spam filters (miscategorizing real email loses valuable messages), churn outreach (calling customers who weren't going to churn wastes CS time).

Why it matters: Precision is the right primary metric when false positives are expensive. A fraud model with 95% precision means only 5% of the transactions it flags are false alarms — low enough that the investigation team's time is well-spent. The tradeoff: higher precision usually means lower recall.

Common mistakes: Confusing precision with accuracy. Accuracy measures all correct predictions against total predictions. Precision measures only the quality of positive predictions. A model can have high accuracy and low precision if true negatives dominate the dataset.

Related recall, F1 score, confusion matrix, AUC, logistic regression