How to read this page — every prediction has two labels: what it learned from, and whether it was made before kickoff or after the game (a backtest). Tap for the full key.
- made before kickoff
- A real forecast: the number existed before the game started.
- BACKTEST: made after the game
- A test, not a forecast. The model was trained only on older seasons and then asked about games that had already been played. It shows how the method would have done — it was never available before those games.
- trained on 2023 and earlier
- The newest season the model was allowed to learn from.
- S, SB, SQ, SBQ
- Versions of the same model that see different evidence. S = plays only; B adds team decisions (depth chart, injury status); Q adds news and talk (beat writers, press conferences).
- v0.1
- The 2026 game forecast the system publishes. It has no fixed training cutoff: it updates on every final score as the season goes.
- Market line
- What the betting market expected, as the middle value across sportsbooks at the last time we saw the line before kickoff. It is the bar to beat. For 2024–2025 we only have a snapshot every day or two, so that line is usually about a day before kickoff, not the final closing line.
- Average miss
- How many points off a prediction was, on average, ignoring whether it was too high or too low. Lower is better.
- 95% range
- If the season had gone a bit differently, the number would very likely land inside this range. A wide range means there are too few games to be sure. (Worked out by re-drawing whole weeks at random 2,000 times. With fewer than 5 weeks of games there is no range.)
- Picked the winner
- Share of games where the team predicted to win did win. Ties and exact pick’ems are left out.
How accurate
Games
Tap a game to see each player's predicted and actual fantasy points.