The track record
Every prediction gets graded against the final score — the good calls and the bad ones. Nothing is deleted, nothing is cherry-picked.
The last three weeks, pick by pick
One square per game, oldest to newest. Tap any square for the call and the final score.
Loading…
The record includes the model's full backtest: for every past game the model was trained only on games before it, then its prediction was graded — the same rules it plays by every night now.
The win chances themselves are calibrated: the model's predicted margin is blended with the Elo rating's probability, with the blend fit only on already-graded past games. When we say 60%, teams in that spot have historically won about 60% of the time. Calibration changes the honesty of the percentages, not the picks — the picked side still comes from the predicted score.