Track Record
How this model has actually done, checked only against games it never trained on. Not a feeling about any one pick. A record.
Is "70%" Actually 70%?
Every game here was held out of training entirely, so this checks whether a stated confidence matches how often the model was actually right, not whether any single pick landed.
200 games · gap +1.4pp
246 games · gap -2.1pp
238 games · gap -1.2pp
170 games · gap +1.5pp
By Season
Predicted Score vs. the Market
Scored the same way, on the same held-out 2025 games the market's own closing lines are compared against here.
Margin
10.2 pts
model · 9.7 pts market
Total
10.7 pts
model · 10.3 pts market
The model doesn't beat the market yet on either number. Shown here plainly rather than only in a JSON file, since an honest "not yet" is still worth being able to check.
Only seasons the model never trained on count above (2023–2025 for the win model, 2025 for the score model). A hit rate on training-season games would be flattering and meaningless: the model already learned from those outcomes.