19_tanking_buylow.md

19 — Tanking Detection / Buy-Low Board (NSBA4 draft)

19 — Tanking Detection / Buy-Low Board (NSBA4 draft)

Goal. Find players whose real (game / external) ability far exceeds what their combine score implies — i.e. tankers / sandbaggers who are underpriced going into the NSBA4 snake draft. Quantify the size of each mispricing. Flag the reverse — combine-inflated players (likely bikers) to fade. Exclude confirmed cheats and staff-tier non-draftables.

• Env: /home/david/code/nsba/.venv/bin/python
• Inputs: player_season_master.csv (game value, standard-SB scoring), combine_ability.csv (de-biked IRT combine ability), team_season.csv, discord_player_signals.csv (qualitative tank flags), docs/INTEL.md, canonical_players.csv (identity).
• Outputs: outputs/buylow_board.csv, outputs/buylow_board.png, outputs/buylow_scatter.png.

Two scoring worlds, never mixed. Game value = standard Science-Bowl PPTF (toss_points / tossups_faced, +4/−4). Combine = pyramidal, gameable tryout (theta/raw from the IRT model in finding 10). The combine is the price; game data is the true value. Mispricing = value − price. Per prior D7 we trust game data over combine.

Method

The mispricing definition

For every player we compute, on a within-pool z-scale:

mispricing  =  z(demonstrated ability)  −  z(combine-implied ability)

Positive = ability beats the combine signal → buy-low (tank candidate). Negative = combine beats demonstrated ability → fade (combine-inflated / biker).

Two tracks, because the NSBA4 draft pool splits cleanly:

• Track A — returning players (18 of the 52 nsba4 combine entrants have prior NSBA1–3 game history). Here both terms are real numbers. z(value) = z of career game PPTF (toss-up points / tossups-faced, pooled over all their clean games, all seasons); z(price) = z of their nsba4 combine theta_overall (the raw IRT ability — see correction below). The two correlate r = 0.74 across these 18 — so the residual off that line is a genuine, mostly-orthogonal mispricing signal. Each player carries a confidence = sqrt(career tossups-faced / max), so a 27-game Vishnu read is trusted far more than a 1-game Ethan read.

• Track B — nsba4-only pool with a Discord tank flag (no game data). The IRT model can't be residualized (no game anchor), so these rest on qualitative intel only: z(price) = nsba4 combine theta; z(value) = a conservative expectation prior from discord_player_signals.tanking_risk (med-tank returning vet → +0.8z, low-tank → +0.3z). Confidence is fixed low (0.40) — these are intel reads, not measured value. This is where the named INTEL tanks (Ziang, nocombomomento) live.

Exclusions (hard)

• Cheat: santhosh_b. (Santhosh V, nsba4 raw ≈ 15) — staff caught him **biking
• leaking answers. His combine is inflated, not tanked**; discount entirely. Removed from the board, listed in the EXCLUDE block.
• Staff / non-draftable: eagle_student, a.new.rag, auride0, jhuang25, xpoes (our GM), thecryolite, isobowldev. None of these appear in the nsba4 combine pool anyway — they self-exclude — but they are listed for completeness.

Results

Buy-low board — top tank candidates (mispricing > 0)

The headline buys. Both top Track-A and Track-B signals agree on the same story.

• Riyan N (nocombomomento) & Ziang Z (deasert_willow) are the two largest mispricings on the board, and they are the two CS-leaning players the INTEL flags as tanks. Both posted near-bottom combines (raw 6–7, ranks 43–45 of 52) yet are returning officers/veterans, both self-describe as CS-capable in a pool where CS is the scarcest, hardest-to-fake category ("CS is unbikable"). The combine is understating them by ~2 z. These are the marquee buy-lows — but on intel only (conf 0.40); they have no NSBA game tape to confirm. Ziang is also a likely captain, which complicates "buying" him.
• Mihir K (modernnewton) is the strongest game-backed tank: a med tank flag, a bottom-of-pool combine (raw 4, 48th of 52), but 13 real games of NSBA play at a respectable 0.24 PPTF. Intel: "college made me know nothing except bio," i.e. rusty/disengaged on the tryout but with real floor. Highest-confidence buy-low (0.71).
• Rohan G — no tank flag, just a quiet undervalue: 16 games, 0.56 career PPTF (top-tier real production, peaked 0.88 in a season) but only a mid combine (raw 22, 10th). High confidence (0.79). The kind of player a vibe-drafting field lets slide.
• Akhil B (abat48) is a special case: he is both a top-3 combine (raw 35) and a top game producer (0.67 PPTF), so his mispricing is small-positive — the combine roughly fairly prices an already-elite player. Not a steal, but not a trap.

Fade board — combine-inflated, likely bikers (mispricing < 0, game-backed)

These have real game data that undershoots their combine — pay the combine price and you overpay. Sorted most-overpriced first:

Caveat on the fades. Negative mispricing means the combine over-states relative to game PPTF, not that the player is bad. Vishnu (−1.20) and Kian (−0.76) are the cleanest "don't pay the combine sticker" flags: both are top-4 combines whose real game tape, while good, doesn't match a top-4 billing — exactly the biking-inflation pattern (D7). Edwin (draoethar) is the cautionary fade: INTEL calls him "the goat" who got better after college and biked chem on his combine; his combine (5th) modestly out-runs his measured nsba2/3 game rate, so he prices high but is plausibly still good — treat his −0.38 as "fairly-to-slightly-over priced," not a true fade.

Cross-checks against the named INTEL cases

• Ziang / deasert_willow → surfaced #2 buy-low (largest game-less mispricing). ✓
• santhosh_b. → excluded as cheat (combine inflated, not tanked). ✓
• nocombomomento → #1 buy-low (the "CS is unbikable" tell + bottom combine). ✓
• Edwin (draoethar) → INTEL "returning star," but his game tape mildly underruns his combine, so he lands on the fade side (−0.38, low magnitude). The model and the hype disagree slightly — flagged, not resolved.

Limitations / caveats

1. Very small n. Track A is 18 players; the buy-low > 0 set is 11. Track-A game reads range from 1 game (Ethan W) to 27 (Vishnu) — read mispricing with the confidence column, which is sqrt(career-TUH/max). The 1–4 game reads (Ethan, Kensuke, Charles, Chris, Edward, Shannon) are barely better than noise.
2. Track B is intel, not measurement. Ziang and nocombomomento's #1–#2 ranks rest entirely on Discord chatter ("genuinely sandbagging," "cs is unbikable") plus a low combine — there is no NSBA game tape for them (conf fixed 0.40). The direction is well-motivated by D7 (low combine + returning officer + unbikable CS), but the magnitude (+1.9–2.0) is a placed prior, not data. Don't over-trust the exact size.
3. Within-pool z-scaling. Both terms are standardized within the nsba4-returning pool (Track A) / full nsba4 pool (Track B), so mispricing is a relative draft-board quantity, not an absolute points number. Cross-season combine→game commensurability is itself only "roughly" established (finding 10).
4. PPTF for nsba3 is a rate proxy (no per-player TUH in the paired layout — see data dictionary); the denominator is estimated for nsba3-only careers. Affects most Track-A players, who are nsba3-heavy.
5. Eligibility / captaincy not modeled. Several top buys are likely captains (Ziang, Akhil per INTEL) — a captain can't be "bought low." Only 5 of 52 nsba4 combine entrants are currently paid & eligible; many buy-low names are not yet confirmed draft-eligible. This board prices ability vs combine; cross it with the eligibility/captain list before acting.
6. De-biking is partial (finding 10): we never see MC-vs-short-answer per question, so theta only down-weights the gameable signal. The fade flags (esp. Vishnu, Kian) lean on the same partial correction.
7. CORRECTION (2026-05-30, red-team M2): the Track-A price anchor is theta_overall (raw IRT), not the de-biked column. Confirmed by recomputation on the 18 returning nsba4 entrants: corr(career PPTF, theta_overall) = 0.740 (the cited anchor), corr(career PPTF, debiked_overall) = 0.669, corr(career PPTF, raw_overall) = 0.759. The 0.74 in this finding is therefore the raw-theta correlation; the "de-biked" framing for Track A was mislabeled. Implication: if you re-price Track A on the de-biked column the anchor drops to r=0.67 and every residual (mispricing) shifts modestly — the ordering of the buy-low/fade boards is unlikely to flip materially (raw and de-biked theta correlate highly for this 18-player nsba4 pool), but the exact mispricing magnitudes are anchor-dependent and should be read as ordinal. See finding 15b for the verification.

Reproduce

Inputs are the just-built combine_ability.csv + player_season_master.csv. The board is outputs/buylow_board.csv (32 rows: 11 buy-low, 13 fade, 8 EXCLUDE); plots outputs/buylow_board.png (ranked mispricing) and outputs/buylow_scatter.png (combine vs game for the 18 returning players, with the fair-value diagonal).

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