One Model. Its Real Record.
This page shows exactly how the Arctic Odds pro football model works, how it was tested, and how far behind the market it still is. If a model seller won't show you this table, ask why.
How It Works
Team Ratings
A recency-weighted ridge regression on point margins. Every game moves every team's rating, with recent games counting more. Regularization keeps one blowout from rewriting a season.
Home Field
A single shared home-field term worth +2.3 points, fit from the data rather than assumed. One number, applied to every home team.
Starting QBs
Starting-quarterback terms fit from every game since 1999. When a starter sits, the line moves by the difference between the two QBs' values — the single biggest lever in football.
The Table Nobody Shows You
Walk-forward backtest, 2023-2025: 816 games, no lookahead. The model only ever saw data available before each kickoff, and the seasons used for tuning never touched the seasons used for evaluation.
Read it plainly: the closing line is more accurate than this model, and 49.9% ATS does not clear the 52.4% a bettor needs to beat the vig. This model does not beat Vegas. Anyone whose backtest says otherwise should show you a table like this one — walk-forward, all games, graded against the close.
Version History, Rejections Included
Teams-only baseline
Recency-weighted ridge regression on point margins with team ratings and a shared home-field term. No player information at all.
+0.69 pts behind the close
+ Starting quarterback
Adds starting-QB terms fit from every game since 1999, so a backup under center moves the line the way it should. Adopted as the live model.
+0.54 pts behind the close
EPA play-by-play models
Replaced point margins with EPA-based play-by-play features. Looked promising in-sample, failed to improve out-of-sample. Rejected.
Rejected out-of-sample
Rest, bye, divisional features
Added rest days, bye weeks, and divisional-game features. No reliable gain in tuning, so they never touched the evaluation seasons. Rejected.
Rejected in tuning
v2 and v3 stay on this page permanently. Documenting what failed is how you know the number next to what shipped is real.
Why Keep Going
A gap of about half a point per game to the closing line is what a clean public model gets you. The market's remaining edge is not math — it is information: injuries, lineups, weather, the things beat reporters know an hour before the model does.
So the roadmap is simple: buy information, publish everything, and grade every week against the close. Each week's lines go up before kickoff, each week's grades go up after, and this table updates whether it flatters the model or not.
If a model seller won't show you this table, ask why.
See This Week's Lines
Every game, the model's line, the assumed starters, and the honest record — free and public.
View the Board