Post-Week-4 player ratings -> subject-resolved tossup-contest simulation on each team's most-recent actual lineup -> home-court category ban -> Normal margin/total. Model-only unless a live per-game moneyline is open (then blended 70/30); the conference-winner FUTURES fallback was DROPPED after the Round-3 backtest showed it hurt (model-only beat the blend on direction 83% vs 50%). Spreads on odd integers, totals on .5 (never push).
Week 6
ChessFun@czz
| ChessFun | czz (home) |
|---|
| Win prob | 57% | 43% |
| Spread | ChessFun -9 |
| Total (O/U) | 158.5 |
home most-likely bans Biology
Week 6
sumin@mingle/Yunyi
| sumin | mingle/Yunyi (home) |
|---|
| Win prob | 54% | 46% |
| Spread | sumin -5 |
| Total (O/U) | 163.5 |
home most-likely bans Physics
Week 6
Fez_Keyreb@anurag
| Fez_Keyreb | anurag (home) |
|---|
| Win prob | 49% | 51% |
| Spread | anurag -1 |
| Total (O/U) | 159.5 |
home most-likely bans Physics
Week 6
dan.k.memes@GidTheKid2
| dan.k.memes | GidTheKid2 (home) |
|---|
| Win prob | 47% | 53% |
| Spread | GidTheKid2 -3 |
| Total (O/U) | 158.5 |
home most-likely bans Earth/Space
Week 6
(Stitch)^2@James W
| (Stitch)^2 | James W (home) |
|---|
| Win prob | 40% | 60% |
| Spread | James W -15 |
| Total (O/U) | 160.5 |
home most-likely bans Physics
Week 6
jonathan@Connor Chang
| jonathan | Connor Chang (home) |
|---|
| Win prob | 40% | 60% |
| Spread | Connor Chang -13 |
| Total (O/U) | 162.5 |
home most-likely bans Biology
Method
Post-Week-4 player ratings -> subject-resolved tossup-contest simulation on each team's most-recent actual lineup -> home-court category ban -> Normal margin/total. Model-only unless a live per-game moneyline is open (then blended 70/30); the conference-winner FUTURES fallback was DROPPED after the Round-3 backtest showed it hurt (model-only beat the blend on direction 83% vs 50%). Home removes one tossup of the category that most helps it; soft softmax over bans (lambda=1, a prior until bans are observed).
Caveats
- LOW CONFIDENCE. Over Weeks 2-3 the model's game-level direction was ~58% (n=12) -- barely better than a coin flip. These are a sharpened prior, not a proven edge.
- Each team is projected with its MOST RECENT actual lineup (who dressed last game), not the full drafted roster -- captures no-shows. A per-player availability model was built + backtested (predicts who dresses at 86%) but did NOT improve the lines (it's star-dominated; see nsba4_attendance_backtest.py), so it stays OFF. Known lineup change? set MANUAL_LINEUP and re-run.
- Model-only this week (no per-game moneylines open yet). The conf-winner futures blend was dropped: on Round 3 it dragged model picks toward season favorites (czz, anurag) who then lost.
- 5 weeks of game data; ratings still lean partly on the combine+draft seed. Byes this week: cryo, xpoes (uncs).
- Real NSBA margins are huge (avg ~45) and hard to predict directionally, so spreads are kept modest on purpose -- widening a coin-flip just amplifies wrong picks. The gain (2.5x) only avoids pick'em.
- Totals are the raw model (no bump): the Weeks 2-3 over-lean reversed (Weeks 3-4 avg ~152, Week 5 went 4-2 UNDER), and the season average is 159.5, so totals now center there. They still lack matchup dispersion (all land ~158-162) -- can't yet tell a shootout from a grind.