The record
Every prediction EdgeBoard has published before a game, and how it turned out. Counted fresh each time this page is opened.
Did 70% mean 70%?
Each point is a group of backtest games where the model gave its favorite about the same chance. The dashed line is where a perfectly calibrated model would sit. The gray bars show how far luck alone can move a group of that size.
Published predictions are drawn as green diamonds on the same axes, and never joined to the backtest line. Until there are a few hundred of them, they will sit all over the chart. That is what a short record looks like.
The explorer splits the record by league, season and month, lists it game by game, and (with Pro) downloads it as CSV.
How to read these numbers
Three scores, two baselines, one rule: a model only earns attention when it beats both baselines on the same games.
Accuracy
How often the model’s favorite won. Easy to read, and the least informative: it cannot tell a 51% call from a 95% one.
Brier score
The average squared gap between the probability and what happened. Lower is better. A coin flip scores 0.250.
Log loss
Like the Brier score, but confident misses cost far more. Lower is better. A coin flip scores 0.693. The settings were tuned to minimize this.
Always pick the home team
Home teams win more than half their games in every league, so this baseline is harder to beat than it sounds.
Decided, pending, void
A published prediction is decided when its game is final. If the game is called off or moved by more than 36 hours, the row is void: still listed, never counted.
Backtest
The same math run over past seasons after the fact. Kept in its own section, with its own label, on every screen.
Check it yourself.
The explorer lists every game, with the number that was published and the result.