<<<<<<< Updated upstream

What the AI predicts

• Spreads — every game, both sides, with confidence on each.
• Totals — over/under on the game total, calibrated against the model's pace + efficiency forecasts.
• Moneylines — straight-up winner with implied probability.
• Player props — top markets (points, rebounds, assists) where the model has enough player-level signal.
• Live in-game — win probability updates per possession, with the same calibration applied to live state.

What makes NBA different

NBA is one of the more efficient prediction targets out of the nine sports we cover, for three reasons:

1. Volume. 1,230 regular-season games + playoffs = ~1,300 games a season. Big training samples mean tight calibration.
2. Player data depth. Box scores, advanced stats, lineup combinations, on-court chemistry — basketball has more granular per-player data than any other sport.
3. Pace + variance. Possessions per game are high, so individual-game outcomes are noisier than soccer or NFL. This actually helps calibration — the model can clearly distinguish high-edge from coin-flip games.

The trade-off: NBA player props are competitive markets. Operators invest more in NBA prop lines than almost any other sport. Pick1 only flags player props where the model sees clear forecast edge against the closing prop line — most nights, that's a small subset.

How accuracy looks

30-day rolling, NBA-only:

Confidence band	Hit rate	Avg CLV
50–60%	54.9%	+0.8%
60–70%	64.5%	+2.1%
70–80%	74.2%	+3.6%
80%+	83.1%	+5.4%

Calibration is tight — the realized hit rate per band lands within ~1% of the stated confidence. For a deeper explanation of how the model is calibrated, see the methodology page.

Common questions

Does the AI predict playoff games?

Yes. Playoff games are in the same model. The training data includes every postseason game since 2018, with playoff-specific features (rest days, series score, home/road game in series). The model tends to be slightly more conservative on playoff confidence because the sample is smaller per matchup.

What about Summer League / preseason?

We skip Summer League — too much lineup variance, players going half-speed. Preseason is generated but flagged as low-confidence by default; most NBA preseason games come in below the 60% confidence band.

Can the model handle injury news?

Yes. The model re-runs whenever an injury report changes. If a star sits out 30 minutes before tipoff, you'll see the prediction recalibrate. Live in-game state (foul trouble, ejections) feeds into the live win-probability head.

Does Pick1 do multi-event combos?

Single-pick predictions only. The math on multi-event combos is unfavorable regardless of edge — combining multiple predictions compounds the market margin. We publish standalone game forecasts; how you use them is up to you.

=======