30_undervaluation.md

30 — Undervaluation-vs-Field Metric (value over the market)

30 — Undervaluation-vs-Field Metric (value over the market)

Date: 2026-05-30. Status: Task C. Script: scripts/A30_undervaluation.py (reproducible). Outputs: outputs/undervaluation_board.csv (241 rows: 189 historical picks + 52 nsba4 combine pool). Env: /home/david/code/nsba/.venv/bin/python.

LEAD CAVEATS (read first). 1. Tiny, leaky sample. Validation rests on 94 drafted players with both a combine and a realized same-season game value, spread over 3 draft seasons (nsba1 contributes ~6 usable, nsba2 46, nsba3 42). Realized value is same-season PPTF — the players were drafted before the games we score them on, but our value model partly reuses those games (the survivorship / same-season leakage that bounds F8). Read magnitudes as directional. 2. This is a re-expression of F8, not a new effect. The metric formalizes the established result that the field anchors on raw combine (|ρ|~0.7) while combine predicts value only ~0.4. The undervaluation score is a systematic, per-player version of that gap. It does not discover a new edge; it operationalizes the known one. 3. No individual name is statistically significant. BH-FDR(0.10) on the 94 per-pick residuals flags 0 individually significant steals/busts. The named boards are shortlisting flags, not significance claims (extends F8 caveat 7, F12). 4. The nsba4 forward board is 100% combine — as gameable and noisy as the combine itself (F3 selection-bias caveat). It predicts who the field will misrank, which is well-grounded, but not that our projection is correct in absolute terms.

The metric

For every player-season we put our value and the field's implied value on a common within-season percentile scale (0–1, higher = better), then subtract:

value_over_field (uvf) = our_value_percentile  −  field_value_percentile

Positive = we rank the player above where the market did → undervalued.

Two field proxies (kept separate on purpose):

Our value = realized game PPTF percentile (F2 north-star) historically; de-biked combine theta percentile for the combine-only nsba4 pool.

Percentile (rank) scale is deliberate: it is unit-free, robust to the nsba3 estimated-TUH proxy and to combine scale drift (F7), and it matches how a draft actually works (relative ordering, not absolute points).

Validation — do combine-undervalued players actually beat their draft slot? YES

The honest test: uvf_combine uses the combine as the field signal; val_resid (F8's steal score) measures beating the ADP/pick-slot regression. These are two different field proxies, so a correlation between them is genuine cross-validation, not an identity.

• Pooled Spearman ρ(uvf_combine, val_resid) = +0.599 (n=94, p<1e-9), bootstrap 95% CI [+0.42, +0.73] (season-clustered resample) — excludes 0. Holds every usable season: nsba2 ρ=+0.53 (p=2e-4), nsba3 ρ=+0.67 (p<1e-5).
• Split test: combine-undervalued picks (uvf_combine>0) beat their slot by +0.088 PPTF-residual vs −0.070 for overvalued picks (Mann-Whitney p<1e-4).
• Robustness (win-shares as our value instead of PPTF): ρ = +0.45 (p<1e-4) — weaker (win-shares is schedule-confounded, redteam T6) but same sign and significant.

Takeaway: players the de-biked/value view ranks above their raw-combine slot systematically out-produce that slot. This is the F8 mechanism made per-player.

Biggest historical mispricings (flags, not significance)

Steals — combine-underrated, then produced (full list in board, realized_src='same'):

Busts — field over-valued (early pick / high combine), under-produced:

Signature is exactly F8's: steals are late picks the combine under-rated; busts are early/high-combine names that under-delivered (Rohan Dhillon #11→neg, owen fei #6). Mihir Kulkarni reappears as a steal, independently confirming F12's buy-low call.

Archetype / positional scarcity — is any type systematically under-drafted?

Mean uvf_combine (we rank above field) and mean val_resid (beat slot) by archetype (F28 labels):

Two honest, partly opposing reads:

1. Bio specialists are the most combine-under-drafted type (+0.159) — the field's raw combine systematically ranks them below their game value. BUT their mean val_resid is only +0.015: they beat combine rank but barely beat their actual draft slot, because the field's live scouting already partly corrects for it. So "bio is underdrafted" is a weak, combine-only effect, not a reliable slot-beating edge. Do not over-weight.
2. The elite multi-science tier delivers the real residual value (+0.237 val_resid, the F28 "this is where wins come from" cluster) even though it is only mildly underpriced on combine (+0.056) — i.e. the market roughly prices the elites correctly on combine, and they still over-produce because combine understates how much the top tier matters. Conclusion: scarcity-adjust toward the broad-elite tier, not by reaching for a specialist archetype. Replacement-tier is correctly faded (−0.117 / −0.084).

This is consistent with the spine (F2/F28): draft for production rate / the broad-elite tier; single-subject specialists (incl. CS, which F28 shows has no standalone archetype) are interchangeable coverage, not a scarcity play.

FORWARD — NSBA4 combine-only board (who the field will misrank)

The field will draft nsba4 by raw combine rank (F8, |ρ|~0.7). We rank by de-biked theta (the bias-corrected ability, F3/F10). uvf_combine = theta percentile − raw-combine percentile. Positive = de-biking lifts them above where the raw board will slot them. 6 meaningful buys (uvf>+0.10), 7 meaningful fades (uvf<−0.10).

Buys (de-biked ability > raw-combine slot):

Fades (raw combine inflates them above de-biked ability):

• Santhosh V lands on the fade list (combine inflated, de-biked low) — consistent with F12's santhosh_b. confirmed-biker exclusion. The metric independently flags him without being told. Treat as confirmation of the fade direction; he should be excluded from any draft entirely per F12, not merely "faded."
• Sanjay O is the most actionable buy: the only elite-multi-science archetype among the top buys, and de-biking lifts him into the upper third.
• Several fades (Rohan G, Roshan A) are high raw-combine names round-1 talk will inflate — exactly where F8 says the field leaks value.

How to use it (and how NOT to): the nsba4 board is 100% combine-derived, so it predicts the field's misranking on solid ground (F8) but cannot confirm our absolute projection (F3 selection bias — combine-high no-shows/tankers break the link, and they are unobserved here). Use it to re-order within a tier, concentrated in the middle rounds (F8: round 1 is efficient), not to invent a tier. Cross every name against eligibility, the captain/keeper list, and the F10 availability/reliability prior before acting — a "buy" who never fields is worth zero.

Limitations / threats

1. 94-row validation, 3 seasons, same-season leakage. The ρ=0.60 is inflated by survivorship (only drafted and ≥2-game players appear) — the true ex-ante edge is the F8 ~13 toss-points/slot figure, of which this is the per-player decomposition, not an independent confirmation of a larger number.
2. Percentile metric compresses extremes. A player at the 0.95 vs 0.99 percentile looks similar; the board ranks order, not gaps. Pair with the PPTF/VORP point estimates (F13) for magnitude.
3. Archetype labels are soft (F28 silhouette 0.23, bootstrap ARI 0.84) — the "bio underdrafted" and per-archetype numbers can reshuffle on a different sample. Reported as tendencies.
4. No multiplicity-corrected individual significance (BH-FDR: 0 of 94 picks). Named players are flags.
5. nsba4 fades/buys assume the field anchors on raw combine as in nsba1–3. If the field de-bikes this year, the edge shrinks. The metric is an exploit of a behavior, not a law.

Reproduce

/home/david/code/nsba/.venv/bin/python scripts/A30_undervaluation.py

outputs/undervaluation_board.csv schema

pool ∈ {historical, nsba4_forward}; season, overall_pick, round, player_raw, canonical_id, discord_tag; raw_overall, raw_rank (combine); realized_pptf, realized_ws, realized_src; field proxies field_pct_adp, field_pct_combine; our value our_pct_pptf (hist) / our_pct_theta (nsba4); uvf_adp (vs draft slot), uvf_combine (vs combine anchor — the headline undervaluation score); val_resid, steal_score (F8 slot residual); theta_overall, debiked_overall, theta_cs; archetype_label. Sort by uvf_combine desc for buys, asc for fades.

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