<<<<<<< Updated upstream

What the AI predicts

• Moneylines — straight-up winner with implied probability. The most useful market in baseball given how often run lines push.
• Run lines — typically -1.5/+1.5 with a price adjustment on each side.
• Totals — over/under with ballpark and weather adjustments.
• F5 (first five innings) — useful for isolating starting-pitcher value before bullpen variance kicks in.
• Live in-game — inning-by-inning win probability incorporating leverage situations.

What makes MLB different

Baseball has more games per season than any other major sport — which is great for training samples, but also means individual-game variance is enormous. A 95-win team and a 67-win team are separated by less than 20% in single-game outcomes.

Three things specifically matter:

1. Starting pitcher. Often the single biggest input. The model has features for fastball velocity trends, breaking-ball whiff rate, recent form, and matchup-specific opponent splits.
2. Ballpark factors. Coors Field plays radically different from Petco Park. Park-specific run environment is a feature on every game.
3. Weather. Wind direction and speed at outdoor parks shifts totals meaningfully. The model uses real-time pre-game weather.

Honest caveat: MLB is one of our higher-variance sports for game-level outcomes. Confidence bands are wider here than in NBA or NFL — the model rarely flags a game above 80% confidence on the moneyline, because baseball just doesn't allow that level of certainty given pitcher and bullpen variance.

How accuracy looks

30-day rolling, MLB-only:

Confidence band	Hit rate	Avg CLV
50–60%	53.6%	+0.5%
60–70%	62.8%	+1.6%
70–80%	71.1%	+2.9%
80%+	78.4%	+4.2%

Common questions

Does the AI predict player props?

Limited. Pitcher strikeout props and total bases for top-tier hitters are supported. Niche batter props are skipped — the per-game sample for any individual batter is too small.

What about the postseason?

Postseason games are flagged with slightly tightened confidence bands because each round is short — a 5-game series is a small sample, and bullpen usage in the playoffs differs systematically from regular season.

Does the model predict no-hitters or perfect games?

No — these are exotic markets where prior probability is so low that any per-game prediction is dominated by base rate. The model focuses on markets where it can produce a meaningfully more informative probability than the public.

Can I use it for fantasy baseball?

Pick1 is designed for game-level predictions, not fantasy player projections. There's some signal overlap (player-level features are in the model) but our outputs are game-outcome forecasts, not lineup-construction-level.

=======