Season
Accuracy
71.6%
11,170 games · the model's pick won
Brier score
0.183
probability error · lower is better
Log loss
0.537
confidence-weighted · lower is better
AUC
0.772
ranking skill · 0.5 = none, 1 = perfect
Calibration
When the model says 70%, it wins about 70% — the probabilities are honest.
The dashed line is perfect calibration. Points on it mean the model's stated probabilities match real-world frequencies — the property that actually matters for a forecasting model, and the hardest to fake.
Reliability diagram · x = predicted home-win probability, y = observed win rate · 10 bins · 11,170 games · ECE 0.028 (mean gap from the diagonal) · both seasons
Versus the market
The closing line is sharp — the model trails it slightly, but stays close.
Model70.2%
Market71.8%
Games w/ odds4,240
Accuracy by how far the model's probability sits from the market's, split by side. Most games have no odds in the data, so treat this as a sample, not the whole picture.
Closing moneylines (sportsdata.io), vig removed · only 4,240 of 11,170 games (38%) carry odds · both seasons
Conviction calls
Where the model disagreed with the market — and was right.
Loading games…
Games where the model and the market picked different winners and the model called it · biggest disagreement first
Every game
The full out-of-sample log.
Loading the full game log…
out-of-sample predictions
showing 0 of 0 games · both seasons · model p = P(home win)