How is draft board accuracy measured?
By rank correlation: how closely a board's pre-draft order matched the order players actually turned out in, scored on classes the model never trained on.
A board is a ranking, so grading one means asking how well its order matched reality's order. That is a rank correlation — Spearman's, specifically. One means the board got the sequence exactly right; zero means it did no better than shuffling.
Over 2018–2022, scored out of sample against Career Score, the Draftanomics board scores 0.52. The analyst consensus scores 0.37. The actual NFL draft order scores 0.60.
Read those three together, because the third is the important one. NFL war rooms rank prospects better than we do, in every year we have scored. They see medicals, interviews and private workouts that no public model has access to. The claim this site makes — the only one it makes — is that it is the most accurate PUBLIC board in the window, ahead of the analyst consensus and behind the teams themselves.
"Out of sample" is doing real work in that sentence. Each class is held out of training before it is scored, so the model is never graded on players it learned from. A model tested on its own training data can be made to look arbitrarily good and predicts nothing.
Measured on drafted players only. Our historical classes are built from draft results, so these figures describe ordering players who were drafted — not the separate question of who gets drafted at all. All three boards are scored on the same players, so the comparison between them is like-for-like.