<<<<<<< Updated upstream

What the AI predicts

• Spreads — every game, both sides, with confidence on each.
• Totals — over/under, with weather and pace-of-play adjustments built in.
• Moneylines — particularly useful in NFL where short spreads make moneylines competitive.
• Player props — quarterback yards, rushing yards, receiving touchdowns where the model has signal.
• Live in-game — drive-by-drive win probability with garbage-time discounting.

What makes NFL different

NFL is the most efficient prediction market on earth. Sportsbooks pour more analytical resources into NFL lines than into anything else. The market closes very tight, and beating the closing line consistently is hard.

Three things help our model in NFL specifically:

1. Injury volatility. NFL injury news creates large, fast line moves. The model re-runs every time the injury report updates and often catches edge in the window between news and full market adjustment.
2. Weather. Outdoor games in cold, wind, or rain see line moves that often under- or over-correct. Weather is a first-class feature in the model.
3. Coaching tendency data. The model has play-call tendency and aggressiveness features per coach — useful for pace-of-play predictions and 4th-down decision modeling.

Honest caveat: NFL is a 17-game regular season per team, which means the per-team training sample is smaller than NBA. Confidence bands tend to be slightly wider in NFL than in NBA — fewer games to calibrate against.

How accuracy looks

30-day rolling, NFL-only (during regular season):

Confidence band	Hit rate	Avg CLV
50–60%	53.8%	+0.4%
60–70%	63.6%	+1.7%
70–80%	72.9%	+3.2%
80%+	81.4%	+4.8%

Common questions

Does the AI predict Super Bowl winners?

Yes — at the season level (futures) and per-game once the matchup is set. Pre-season Super Bowl futures are notoriously hard for any model because the sample of championship outcomes is so small relative to roster turnover.

What about the playoffs?

Same model. Playoff games are flagged with slightly lower confidence default because each round is a single elimination game and variance dominates on shorter samples.

Does the AI predict player props?

Some. The model surfaces top-line props (QB pass yards, RB rush yards, top WR receiving yards, anytime TD scorer) where it has clear edge. Niche props (e.g., longest completion) are skipped — the training signal isn't dense enough.

Can it handle Thursday Night / Monday Night games?

Yes. The model has features for short-rest games (Thursday) and prime-time effects. Both are handled the same way as any other game.

=======