DomModel is back.
After a brief one-year hiatus, DomModel returns for the 2026 NFL season with a new, experimental methodology. Previous iterations of the model sought to maximize closing line value by predicting closing spreads. This version uses similar techniques but predicts realized spreads directly. As a result of the shift, performance data from prior versions are not applicable to this model.
The model will be accessible either via DomModel.bet or via the "DomModel" link at the top of the FirstAndThirty homepage. The site will be updated once a week after Dom has a chance to run the model. You'll be able to see the model's projection for every game, every week, completely free.
These sheets intentionally separate model outputs from handicaps and wager recommendations. The model's focus is solely on quantitative methods for predicting NFL spreads. I (Brady) will review the results each week and post my own recommendations on FirstAndThirty based on the model's output.
Home-field advantage is incorporated into the model. Week-to-week factors such as injuries are more difficult to quantify, but we'll address them where possible in weekly commentary. Gameday factors, including weather and game-time decisions, are even more problematic. The model does not account for hooks.
You may cite or reference this model on a sports betting website, blog, podcast, YouTube channel, or similar outlet with appropriate attribution. Any wager recommendations must be clearly identified as your own, and you may not imply that model outputs constitute a recommendation to place any wager.
Highly Technical Shit for Nerds
The selected threshold is 1 point, chosen through a combination of gut instinct and statistical backtesting. Leave-one-season-out, out-of-fold validation produced 563 bets, 308 wins, and 54.7% accuracy, with a 95% Wilson confidence interval of 50.6%–58.8%. Each held-out season contained at least 83 bets.
Per-season accuracy was:
- 2020: 53.1%
- 2021: 55.4%
- 2022: 56.1%
- 2023: 54.2%
- 2024: 55.0%
The reported confidence interval applies to the selected out-of-fold threshold results and thus treats the 1-point threshold as if fixed in advance. Because the threshold was selected using the same backtest, the interval is mildly overoptimistic and some apparent performance will reflect selection noise.
You are free to apply whatever point threshold you want when using the model in your own handicapping. The model has not been forward-tested, and any practical implementation will still require informed handicapping.
