NSBA Lab · How the Numbers Are Made
Methods & Caveats
What's in the dataset, how every figure is computed, and an honest accounting of what these numbers can and can't tell you.
01The data
the corpus behind every claim
Everything on the Lab is computed from the raw buzz and game logs — no third-party stat feeds, no manual tallies. The scripts are pure Python standard library, and scoring is canonical Science Bowl throughout: tossup +4, neg −4, bonus +10. The corpus spans four leagues:
- NSBA4 — the current draft-fantasy season: 14 teams, 42 games, 6 games each.
- NSBA 1–3 — three prior draft-fantasy seasons: 49 teams, 188 games.
- dasoni-2 — an open scibowl.live tournament of undrafted school teams: 43 teams, 122 games.
- SSB-2026 — three combined open tournaments (Johns Hopkins Invitational + Stanford SB + Stanford College): 84 teams, 285 games.
| League | Type | Teams | Games | Buzzes |
|---|---|---|---|---|
| NSBA4 | draft fantasy | 14 | 42 | 1,331 |
| NSBA 1–3 | draft fantasy | 49 | 188 | — |
| dasoni-2 | open | 43 | 122 | 5,163 |
| SSB-2026 | open | 84 | 285 | 13,632 |
02The honest caveats
what these numbers can and can't tell you
- This is exploratory, not a controlled study. One small dataset, dozens of correlations, and no multiplicity correction. Treat the NSBA (n=14 teams) figures as hypothesis-generating and expect some to be noise.
- NSBA4 is tiny — 14 teams, 6 games each. Its competitive-balance ratio (1.02) carries a wide bootstrap CI [0.67, 1.25]; the measurement-illusion point is that with this little data "equal" and "hidden talent" aren't separable.
- Pro-league Noll-Scully values are cited, not recomputed, and aren't directly comparable to a 6-game season — which is why the analysis leads with a simulation, not the raw comparison.
- Availability, coverage, and star-out effects are observational, with small, confounded cells (a team missing a subject or a player is usually just weaker); read them as associations and directional magnitudes.
- Buzzing has survivorship — we see buzzes taken, never buzzes declined — so accuracy-by-position describes the buzzes players chose, not a "would waiting help" experiment. Celerity isn't comparable across formats.
- Correlation ≠ causation throughout; raw talent confounds nearly everything. dasoni-2 is a scrimmage set (secondary replication); "nationals" is a proxy (Stanford SB, JHU Invitational).
Pressure-tested. The analysis was run against an adversarial panel — a statistician, a scientist, a quant, and a sports GM — and several headline numbers were corrected as a result.
NSBA Lab · built from scibowl.live tournament exports + the NSBA season database · July 2026