Eick Bet is a picks and analysis site. Every game in a week is priced from information available Thursday at noon Eastern, before Thursday night. The margin is an Elo and EPA blend, plus rest, and a prior-week starter who is out or doubtful on a report dated before that cutoff. Recent form is tracked and then left out of the number. Recorded weather is not an input. A side is a pick only when its edge clears 3% and the expected value at the posted odds is positive. Spread probabilities use a wider sigma chosen on 2015. Weeks 1–4 spread picks are labeled Caution and are not called high confidence. The 2016–2025 record is retrospective development. Games from here on are the proof.
Margin
Elo and EPA are each turned into a home-margin forecast and blended 55/45, then adjusted for rest and a prior-week starter who is already ruled out. The schedule's actual starter is not an input. Recent form is not added. The margin is normal. The spread uses the 2015 residual sigma times a scale chosen to calibrate 2015 cover probabilities. That scale is not refit on later seasons.
Total
A shrunk points model and an EPA scoring model are averaged. Recorded stadium weather is not an input. The total is normal, with its own 2015 sigma.
Props
Pass yards, rush yards, receiving yards, and receptions use a shrunk player rate times an opponent factor clamped between 0.75 and 1.25. Yards and receptions use a normal with variance frozen on 2015. Rookies without two prior games are not priced.
Walk-forward, on purpose
Ratings start in 2009. Every feature for a game is computed before that game is allowed to move Elo, EPA, form, or player rates. Games on the same day share a state, so a 1 p.m. result cannot leak into a 4 p.m. projection. Thursday can inform Sunday, because it is a previous day.
The EPA feature is converted to points with a regression fit only on 2009-2014. That fit has an intercept and a home indicator. Neutral-site games are rare, so those two numbers are not a clean split of home field. On an ordinary home game with an even EPA matchup they add to about 2.6 points, which is the figure that matters. The margin and total are recentered, and their sigmas are set, on 2015 only. Seasons 2016-2026 are the test. Weights, the 3% threshold, and the confidence tiers were written down before that test and were not searched against it.
Closing lines are the nflverse schedule prices. Home is favored when spread_line is positive, and home covers when the home margin beats that number. Moneyline and total pushes void the stake. Odds that are missing are filled at -110 and counted. The market is not blended into the projection (market_blend is 0, so the posted line is the price we grade against and not an input to the projection. Raising it pulls the projection toward that same line and shrinks measured edge.)
Player-prop closing lines are not in the free files, so the accuracy page does not invent a prop betting record. It reports whether the projection is biased and whether the interval covers what it claims. A historical file can be dropped in later. See the repository notes.
Adding an input
Weights and thresholds live in config/model.json. A new input is a feature computed in the walk-forward pass from information dated before kickoff, a weight in that file, and a line in the projection. Do not refit the weight on 2016–2026. Re-estimate scale only inside 2009–2015, then rerun the backtest.
Live odds are optional. Set THE_ODDS_API_KEY and rerun the backtest script. The key stays in the environment. The site itself is a static board of the JSON that script writes.