Updated 2026-07-25
Precision@N measures the share of the top N ranked predictions that turn out to be correct. It comes from ranking and search evaluation, and it fits any problem where only a short list gets acted on. If a model ranks candidate levels and a trader looks at the strongest handful, then scoring the whole candidate set is the wrong question. Scoring that handful is the right one.
Two properties make it honest and also demanding. It depends on N, so the cut has to be stated rather than chosen after the fact, and it says nothing about the levels outside the top N, which is what reach rate and recall cover. Reported without a sample size and an interval it is close to meaningless, which is why every figure we publish carries both. We do not quote a number here on purpose: the current measurements, their sample sizes and their Wilson intervals live on the proof page.