2026-06-06 · the full readable workspace

NSBA Draft Analytics

🆕 Latest updates

• Live draft tool — built & tested: a stateful engine for draft day (14-team × 6-round snake). Plug in your slot, feed it picks as they happen, and it recommends by roster-aware marginal value (your open subjects weighted up, depth in your best subject once covered) with a "won't last" slide flag (the field drafts the raw combine, so we predict who's gone before your next pick) and honest floor/ceiling. One-command data pipeline ingests new combine + intros. → Live Draft runbook
• Joint hierarchical model — the big modeling rebuild (honest result): replaced the stitched combine-IRT + game-rate pipeline with one Bayesian model — a per-player 6-subject ability vector, combine + game as two views, biking auto-down-weighted, scibowl difficulty as a prior, full posterior uncertainty. It fixes the specialist dilution (William Wang's chem surfaces, etc.). But out-of-sample it's only combine-grade on raw ranking — n is the wall. Its real value is per-subject fit + honest floor/ceiling, not a predictive miracle. → 43 joint model
• Backtest scoreboard — every model is now falsifiable: a harness that replays the nsba2/3 drafts and scores any ranking by realized win-shares. The bar (combine-greedy) and ceiling (oracle) are fixed; nothing gets trusted until it beats the bar out-of-sample. → 42 backtest · 41 scibowl anchor
• Draft Strategy v3 — we red-teamed our own "generalist bias": David & Anurag pushed that the engine over-favors generalists. They're partly right. PPTF dilutes a specialist's rate ~3× (it charges them for 4 cycling categories teammates cover) — but the bias bites the mid-board, not the elite tier (top-10 unchanged), and the under-priced specialists are bio/ESS, not CS (CS is over-supplied at the top — this refutes "reach for CS"). The D14 playoff rule also rewards depth in your best subject (thin→stripped loses 24% vs 9% with a 2nd body). → 40 strategy v3
• Kian & Vishnu fades, re-litigated (Anurag is half-right): Kian downgraded fade→neutral — a genuine bio/chem specialist & top-25 producer, just not the #1-overall his combine implies (don't pay #1, do take as a 2nd-round specialist). Vishnu fade STANDS (high conf): 27 games of average-plus behind a +20 combine climb, and his self-claimed "CS" is his worst in-game subject (net −8). → 40
• NSBA4 #introductions intel — parsed: the live pool self-reports (40 draftable; 34 combine rows, 12 with game tape). Pool is thin-but-better-than-feared — the marquee names all showed up. Confirms the buy-low board (Rohan G self-fades "washed" = bullish) and starts the eligibility/captain list. Combine-skippers with game tape = the field can't price them (a structural steal). → 39
• Identity-gap fix (the bug David flagged) — done: draft→game link rate jumped nsba1 19%→43%, nsba3 65%→77% (43 same-person IDs that were split across sources, merged via combine-sheet "Rosetta" bridges). A 52-row worklist is ready for David to confirm — start with 4 ID collisions (e.g. "evan" = 5.59 WS, must split into Evan Zhang + Evan L). → 37
• Anurag's "cleanup get" idea — tested, a wash: a get after the opponent negs (15% of all gets) is conceptually "cleaning up, not out-buzzing," but cleanup-rate is uncorrelated with skill (−0.006), isn't a stable trait, and discounting it doesn't move the board (top-20 unchanged) or predict better out-of-sample. Raw PPTF already absorbs it. → 38
• Buzz-rank vs difficulty-weighting — settled on REAL pyramidal quizbowl data: none of the smart buzzpoint features beat opponent-adjusted outcome-only out-of-sample. thedoge's buzz-rank bet actually loses (it throws away conversion volume — buries elite high-volume players); Radius's difficulty-weighting is a wash. The ranking ceiling holds. → 36
• Usage / empty-stats bias — tested & resolved: the "lone star on a bad team gets over-credited" worry is real but small and two-sided; the teammate-adjusted model gains nothing out-of-sample, so raw PPTF is vindicated as the draft currency. → 32 measure · 33 adjusted · 93 skeptic
• Player reports — new section in the sidebar (thedoge42, Anurag Sodhi, arolakiv, Jonathan Huang, Sanjay), plus a /player-report skill to generate a card for anyone.
• NSBA1 draft analysis + win-shares — incl. the Vishwa Reddy Akkati / "Vish" answer (biggest steal of the draft). → 34
• Draft Strategy v2 + Bonus value — bonuses are 55–61% of scoring but carry no per-player signal; strategy unchanged. → 26 strategy · 27 bonus
• Plackett-Luce / buzzpoints test — done: dan's own ranking model uses no buzzpoints at all — the ranking signal is opponent-adjustment, not buzz position (adding buzzpoints moved the scibowl ranking by ρ=0.978, i.e. nothing). Radius's skepticism holds, but the load-bearing ingredient is a model, not raw stats. → 35

The headline (skeptic's version)

• Value = production rate (PPTF). corr with win% ≈ +0.62, survives every control. How points are spread across a roster barely matters.
• The combine is a gameable r≈0.6 seed — and the field drafts it almost mechanically (ρ≈0.7), leaving ~13 toss-points/slot of edge for a value-based GM.
• Coverage is table-stakes the draft auto-provides (teams cover all 6 ~96%); "don't punt" does not survive a design rebuild.
• CS is modest, not 3.5× (~2× on toss-ups, 1.2–1.4× with controls). Secure one good CS body via theta_cs; don't reach.
• Pick value is steeply convex — never trade your R1 down for volume. The NSBA2 1st-for-two-2nds trade was −2.73 win-shares (gave up Daniel Sun, a star; finished 9th while his new team finished 2nd).

Everything is ±a lot (n=41 team-seasons). Trust directions; read the methodology and causal red-teams for where magnitudes are soft. Full detail in the Synthesis.

Charts

Growth / agingCS valueBuy-low boardBuy-low scatterCombine IRT abilityCombine → gameDifficulty factorsnsba1_winshares.png

All documents

NSBA Draft Analytics · embargoed until after the SSB draft · ← hub