The best public forecast, the closing betting line, picks the winner of about 66–67% of NFL games. Good statistical models land between 63% and 67%. Ours picked 65.0% over 4,146 regular-season games from 2010 to 2025 in a walk-forward backtest, and the live record is published every week. Nobody sustains 70%; a single 17-game season swings several points on luck alone.

No. Predictions are published to show whether the method works, and every one is graded in public. We never recommend bets, we do not link to sportsbooks, and our own analysis found no profitable edge against the closing line. Please treat the numbers as a study, not a tip.

Yes. Picks are stored before kickoff, with the probability and a plain-language reason, and the page shows them as soon as they exist, usually a week ahead. The final version is the last one stored before kickoff; the What Moved panel shows how it changed on the way there.

Schedules, results, play-by-play, injury reports and depth charts come from the open-source nflverse project (CC-BY-4.0). Market prices come from Polymarket's public API and from DraftKings lines as shown on ESPN's scoreboard. News comes from public RSS feeds. Weather comes from the National Weather Service. Nothing is paid for and nothing is scraped from behind a login.

Never as an input. The closing line is stored and shown only as the benchmark to grade against. That is deliberate: a model that copies the market cannot tell you anything the market does not.

Because a projection is the middle of a wide range. Real NFL scores vary by about ten points per team from game to game, but the most likely score for most games is somewhere near 21–27. A model that projected wild scores would be less accurate, not more. The win probability carries the real information; the score is a companion.

It agrees with the market on the favorite about 87% of the time, because both are looking at the same teams. When the two disagree, history says the market is right more often (about 56–44). We publish those disagreements rather than hide them; they are the interesting games.

Two things, and we want to be clear about both. First, AI helped build this: the code, the pages and much of the writing were produced with an AI assistant (Anthropic's Claude) working alongside a human owner who set the goals, made the decisions and checked the results. Second, AI runs one step of the weekly process: a language model reads news reports and turns them into structured, capped claims (for example, a quarterback change the schedule data has not caught up with). What AI does not do is make the predictions. Those come from ordinary statistics — regressions on team efficiency, quarterback play, injuries and rest — that we can explain line by line. Every AI-produced claim is logged with its source, limited to a few points, and graded afterwards, so you can see exactly what it changed and whether that helped.

Four times a day, plus extra runs on Sundays before the early, late and night games so inactives are counted. Each run refreshes the data, grades finished games, reads the news, re-predicts, snapshots the markets and rebuilds the rankings and season projections.

Usually because the named starting quarterback changed, the injury report changed, or a news item passed the analyst layer's filters. The What Moved panel on each week's page lists every change with the biggest reason. Nothing changes after kickoff.

It projects a margin and a total for every game, and both are graded against the closing lines on the Record page. Expect those records to sit near 50%: beating the closing spread consistently is something almost no public model does, and we say so.

Not yet. The model rates quarterbacks individually because the starter matters more than any other single player; other individual stats are a separate problem we may study later, and if we do, they will be graded the same way.

niftiNFL is a project of niftiIQ Lab, part of Nifti Mobility Group in Biddeford, Maine. It grew out of an NBA model built the same way, and both are published to show whether a careful, honest method works over time.