NFL Predictions 2026: Every Game, Frozen Before Kickoff, Graded After

This page is a ledger, not a tout sheet. Before each week the model — a fully-stated walk-forward Elo that reads nothing but final scores — publishes a win probability for every game. Those numbers freeze at publication (the JSON is served raw, CSV here), results attach as games finish, and the scoreboard below keeps the season-long grade. It knows nothing about rosters, injuries, or quarterbacks — which matters — and it will be wrong roughly one game in three, because in its 16-season backtest that is exactly how often it was wrong. This is not betting advice; for why beating the closing line is a different and harder claim, see how accurate the spread actually is.

The scoreboard

The ledger opens with 16 frozen predictions (Week 1, locked Aug 12). Grading starts the moment the first game finishes and never stops; every probability on this page was published before kickoff and cannot be edited afterward.

Week 1

Locked Aug 12. Probabilities are the model's, stated from each side's rating (no home edge applied at neutral sites); the favorite shown is simply the side above 50%.

Current ratings, all 32

post-2025, after 33% preseason regression toward 1505 — the scale is relative (average ≈ 1505); a 100-point edge at a neutral site is about 64%.

The rules this page lives by

• Frozen means frozen. A published probability is never edited. Rebuilds can only attach results.
• A missed week stays missed. If a week's numbers weren't published before kickoff, that week shows no predictions — there is no such thing as a retroactive forecast.
• The bad weeks stay up. The season grade includes every game the model got wrong, on the same page, forever.
• Same engine everywhere. The script that generated the backtest is byte-for-byte the one that generates these numbers.

Sources: nflverse public game log bundled at data/games.csv; the model and ledger are generated by the site's build (nfl_elo.py) with no hand-entered numbers anywhere.