Gridiron Oracle
An NFL forecasting model — margin, total, win probability and value against the market — plus fantasy football rankings. Fully public data, strict walk-forward validation, and results reported without spin.
The honest result, in one line
The model matches the market’s closing line. It does not beat it. Across 3,908 out-of-sample games it posts a Brier of 0.2130 against the books’ 0.2122, and a margin MAE of 10.03 against 9.97.
This is exactly what should happen, and it is the best possible news: the NFL closing line is one of the most efficient estimators in any market anywhere. Any project claiming to beat it consistently by a wide margin, on public data, is overfitting or measuring wrong.
Where the real edge is, and why it is not here yet
This backtest is validated against the closing line. Nobody bets the close. Money is made against the opening line, against slow books, and on injury news before the market digests it. That is roadmap work, not a promise that the model already does it.
What it does add
- A discrete distribution with key numbers. NFL margins pile up on 3 and 7; a normal distribution makes systematic errors at exactly the lines where the money is.
- Parameterized on the market residual. The model predicts where the line is wrong, rather than predicting the margin with the line as just another feature.
- Calibration measured and published. For betting, the probability matters more than the ranking.
- A standalone model that never looks at the line, so we can answer whether the on-field signal is worth anything on its own.
Known limitations
- Validated against closing lines, the hardest to beat.
- No live injury data; the QB effect is inferred from the announced starter.
- Weather is historical and observed, never forecast.
- No multi-book lines, so no line shopping.
- No player props or alternate markets.