Panel Ruling: NBA Player Props Data Audit & Custom Metrics

Date: 2026-03-31 Panel: Opus 4.6, Sonnet 4.6, Gemini 3.1 Pro, Grok 3, GPT-4.1 (via OpenRouter) Topic: Audit the NBA player props data collection, identify gaps, and recommend custom metrics Winner: Opus (Advisor E) — 24/25 points

Scoreboard

Consensus (3+ advisors agree)

Missing Data — Must Collect

1. Usage Rate (USG%) — The single most important missing metric. Formula: 100 * ((FGA + 0.44*FTA + TOV) * (Team_Minutes / 5)) / (Minutes * (Team_FGA + 0.44*Team_FTA + Team_TOV)). Source: computable from existing game logs + team box scores.
2. True Shooting % (TS%) — PTS / (2 * (FGA + 0.44 * FTA)). Computable from existing data.
3. Per-36 and Per-100 Possession Rates — Raw material for minutes-adjusted projections. Isolates pace effects.
4. Assist Rate, Rebound Rate — Position-relative capture rates, independent of pace/minutes.
5. On/Off Splits — How Player X's stats change when Teammate Y is on/off court. Source: NBA Stats API playerdashonoffdetails (free).
6. Player Tracking Data — Touches, drives, potential assists, hustle stats. Source: NBA Advanced Stats API (free, rate-limited).
7. Lineup/Stint Data — On/off per 100, minutes in 5-man units. Source: Cleaning the Glass (~$25/mo) or NBA Stats API.

Matchup Adjustments

• Use Pace-Adjusted Defense vs Position (DvP), not raw "team allows X to SFs"
• Group players by offensive role (Primary Ball Handler, Wing Scorer, Spot-up, Rim Roller, Versatile Big), not just position
• Opponent Defense Multiplier with 30-day rolling window
• Shrinkage factor (lambda 0.2-0.4) to blend player-specific and population priors

Custom Metrics to Build

1. Minutes Projection Model — Highest ROI gap. Weighted: 40% L5 + 30% L10 + 20% L20 + 10% season. Factors: B2B, rest days, opponent pace, blowout probability.
2. Pace + Usage Adjusted Projections — Player_Rate_per36 / 36 * Projected_Minutes * (Tonight_Pace / League_Pace) * Usage_Factor
3. B2B/Fatigue Multipliers — Player-specific, not population averages. Average impacts: Points -4.2%, Assists -6.8%, Threes -7.4%, Rebounds -2.1%.
4. Blowout Risk Adjustment — Reduce projected minutes 35-55% for non-stars in projected blowouts (>12 pt lead in 4Q).
5. Variance Profiles — Player-specific over-dispersion parameters. Profile classes: consistent (CV<0.25), volatile, boom-bust (CV>0.5).
6. Correlation Matrix — Pairwise stat correlations per player for PRA/combo/double-double pricing.

Distribution Upgrades

• Zero-Inflated Poisson/NegBin for blocks, steals, and low-volume threes
• Player-specific dispersion via Bayesian shrinkage (not league-wide constants)
• Matchup-dependent variance — inflate sigma vs elite defenders or weak rebounding teams
• PIT calibration — probability integral transform test + Brier scores vs FD implied

Noise to Trim

• Raw plus_minus (extremely noisy)
• Single-game fg3_pct or ft_pct (almost pure noise)
• Season averages without recency weighting
• Generic team off/def rating without pace or role adjustment

Key Unique Insights

Opus (winner): The event-driven snapshot window — the 30-60 minute lag after a star is ruled out is the most exploitable moment in player props. Also: Bayesian shrinkage for player sigma, cleanest usage-shift model.

GPT-4.1: 13-metric audit table, 7-step matchup process with shrinkage tuning, season-stage adjustment (league priors early, player-specific after game 50).

Sonnet: Flagged Normal distribution for points as inappropriate — should be skewed/mixture. Sportsbook-specific line behavior (DK/FD/PrizePicks move differently).

Gemini: Skewness and kurtosis modeling for tail events, Kalshi alt-line specific optimization.

Grok: DFS ownership as contrarian signal, five named custom metrics.

Blind Spots (all 5 missed)

1. Referee assignment data — NBA ref crews have 15-20% variance in foul rates, directly affecting points/assists/rebounds/steals props. Free data, available ~9AM ET game day.
2. Garbage-time stat inflation — Bench players padding stats in blowouts contaminates rolling averages. Need score-differential weighting.
3. Roster transaction lag — Post-trade data reset windows are high-edge but no advisor specified how to handle them.
4. Book-specific limit structures — A +EV prop is useless if the book limits you to $50 after one winning month.

Raw council files: /home/ubuntu/edgeclaw/data/councils/2026-03-31/player-props-audit/