Each row is built from every YES/OVER prediction the model emitted in the last 90 days that has a settled outcome.
Hit rate excludes pushes (stake refunded). A perfectly calibrated model has hit rate ≈ mean modelProb across the bucket. Mean Brier is the average squared error of the YES probability — lower is better, 0.25 is a coin-flip baseline. Realised edge is the average flat-stake return per priced pick: positive means the model beat the book's price after the fact.
42predictions still pending — usually corners / cards (need API-Football key) or NBA player props that we haven't yet stored player IDs for.