Win probability
Game Prediction Model
- 20seasons of data2006–2025
- 5,133games trained on4,166 more held back for testing
- 64.7%called correctlyon games the model never saw
- ±2.1points of calibration driftwhen it says 70%, it happens about 70% of the time
23.2%decided by 3 points or fewer
47.5%decided by 7 points or fewer
8points, the median margin
A quarter of NFL games come down to a field goal, which is the ceiling every forecaster runs into. Predicting every game decided by more than a touchdown perfectly, and coin-flipping the rest, would still only reach 76.3%. Across 5,133 games, 2006–2025.
What the confidence means
Accuracy above roughly 65% is bought by declining to call the close games, not by better modeling. Never quote an accuracy figure without the coverage it applies to.
Each band, and how often it has been right
| Band | Model says | Share of games | Right this often |
|---|---|---|---|
| High confidence | 65.0% or better | 42.1%(1,756) | 74.1% |
| Medium confidence | 60.0% – 65.0% | 19.5%(812) | 63.4% |
| Low confidence | 55.0% – 60.0% | 20.1%(838) | 57.3% |
| Coin flip, no call | 50.0% – 55.0% | 18.2%(760) | 52.8% |
Being choosier buys accuracy, and costs coverage
Each row is the same model, ignoring the games below a threshold. Accuracy rises only because the hard games have been dropped. The modeling does not improve down the table.
- 50.0% or better64.7% on the 100.0% of games it calls
- 55.0% or better67.4% on the 81.8% of games it calls
- 60.0% or better70.7% on the 61.6% of games it calls
- 65.0% or better74.1% on the 42.1% of games it calls
- 70.0% or better77.7% on the 26.0% of games it calls
- 75.0% or better80.7% on the 13.4% of games it calls
- 80.0% or better84.5% on the 5.7% of games it calls
Calibrated Logistic Regression (core + market features) · walk-forward by season (train on all prior seasons, test on the next)