Model Accuracy
We believe in showing our work. Here's exactly how accurate our win probability model has been across every game we've analyzed.
Every percentage on this page is outright accuracy: did the team we favored win. For how that differs from ATS and Brier score, see the accuracy glossary.
Factor model scores every current game. Each model's record is shown separately, never blended into one number.
Overall Performance
Calibration by Confidence
When the model says 70%, does the favored team actually win ~70% of the time? Points above the dashed line outperform; below underperform.
| Model Confidence | Games | Correct | Actual Accuracy | Calibration |
|---|---|---|---|---|
| 50-55% | 12 | 7 | 58.3% | Outperforming |
| 55-60% | 7 | 2 | 28.6% | Underperforming |
| 60-65% | 5 | 3 | 60.0% | Well calibrated |
| 65-70% | 4 | 2 | 50.0% | Underperforming |
| 70-75% | 3 | 3 | 100.0% | Outperforming |
| 75%+ | 1 | 0 | 0.0% | Underperforming |
Week-over-Week Trend
How the model has performed over recent weeks. The dotted line marks 50% (coin flip).
How We Measure Accuracy
A prediction is correct when the team our model gives > 50% win probability actually wins the game. Ties are excluded from accuracy calculations.
Calibration measures whether our confidence levels are meaningful. If we say a team has a 70% chance of winning, they should win roughly 70% of those games -- not 90%, not 50%.
For context, a coin flip gives 50% accuracy. Our model targets outright winners: predicting which team wins the game.
Model-vs-market track record
Games where our model and the betting market disagreed by 5+ points at kickoff, frozen at the start with the odds and probability shown then, and graded after the final score. These are disagreements with the market, not proven value, and past results say nothing certain about the next game.
24 of 100 leans graded
Tracking started on Sep 26, 2026. We publish the record once 100 leans are graded; a handful of games says nothing about a model.
Not betting advice. Responsible gambling