WNBA Research Pipeline — Council Ruling

Date: 2026-04-01 Process: Full 5-phase council (Advisory → Anonymization → Peer Review → Chairman Synthesis → Boss Ruling) Advisors: Opus, Sonnet, Gemini 3.1 Pro, Grok 4.20 Reasoning, gpt-oss-120b Winner: Grok (tied with gpt-oss at 1 genuine vote each; Sonnet endorsement of Grok breaks tie) Status: PENDING BOSS RULING on open questions

COUNCIL SUMMARY

Where Advisors Agreed

1. Player availability is #1 variable — smaller rosters (12 players) mean single star absence moves every market
2. 40-game season creates data sparsity — stats don't stabilize until ~15-20 games, need Bayesian priors from prior seasons
3. Olympic/national team absences are unique WNBA risk — mid-season multi-week absences not seen in other leagues
4. 7 edge scanners needed — Moneyline, Spread, Totals, Series, MVP, ROY, Draft #1
5. Pinnacle coverage may be thin — wider vig than NBA, may need synthetic sharp line from analyst consensus
6. Expansion teams create model gaps — no historical data, need rapid calibration approach
7. Overseas league performance is intelligence source — offseason leagues inform player form entering WNBA season
8. Commissioner's Cup changes team motivation — mid-season tournament alters lineup priorities
9. Charter flights (new) change rest/travel dynamics — historical rest data may not apply
10. Coaching changes have outsized impact — only 12 teams, coach quality variance is high

Where Advisors Disagreed

1. Model approach for small samples: Some proposed standard regression, others Bayesian shrinkage with NBA priors. Council verdict: Bayesian with WNBA-specific priors from prior seasons, with NBA analogues only as last resort.
2. MVP/ROY modeling: Some treated as pure stats projection, others included narrative/media voting patterns. Council verdict: Hybrid — stats foundation + media narrative tracking + voting history analysis.
3. Draft market approach: Some proposed mock draft aggregation, others workout intelligence. Council verdict: Mock draft consensus + GM signaling + workout report tracking.
4. Expansion team handling: Varying approaches from ignoring to imputing. Council verdict: Use roster-based projection from player history, wide confidence intervals.

Strongest Arguments (from peer review)

Grok wins (endorsed by Sonnet) with the best judgment-over-execution approach:

• Correctly identified that WNBA's thin market means discipline matters more than sophistication
• Focused on where genuine edges exist (availability speed, market inefficiency) vs where they don't
• Honest about model limitations with 40-game season
• Practical kill switches for expansion team volatility

Opus strong runner-up (self-voted but Opus review insights were deepest):

• Concrete edge thresholds per market with reasoning (4% ML, 5% spread, 10% ROY, 12% Draft)
• Three desk-killing risks identified with mitigations
• Paper-trading mode for first 3 weeks before going live
• Synthetic sharp line when Pinnacle coverage absent

Biggest Blind Spot

Correlated exposure across markets within same game — All advisors designed 7 independent scanners but none addressed that ML, spread, and totals for the same game are ~1.5-2 independent bets, not 3. In a 12-team league where one star absence moves all markets simultaneously, treating each scanner independently is concentrated risk. Need correlation matrix between scanner outputs and portfolio-level position sizing.

What Everyone Missed (from peer reviews)

1. Late-scratch arbitrage windows — Time delta between pipeline detecting a scratch and Kalshi repricing is 5-15 minutes (vs seconds in NBA). Pre-built conditional orders on scratches = pure speed edge independent of model accuracy.
2. Regime-aware backtesting — 40-game seasons with expansion, charter flights, and rule changes create structural breaks. Standard backtesting across seasons is misleading.
3. Market execution latency — WNBA Kalshi contracts are event-based (not continuously market-made), meaning order flow is thin and pricing lags are exploitable.
4. Within-game market correlation — If model finds ML edge, it almost certainly finds spread and totals edges too. Not three bets — closer to 1.5.

BUILD PLAN

Phase 1: Core Data Tables

wnba_teams: team_id, name, abbreviation, conference, arena, charter_flight (boolean), expansion_year, coach_id, active wnba_players: player_id, name, team_id, position, height, age, experience, salary, overseas_team, national_team, star_tier (1-3), active wnba_games: game_id, season, date, home_team, away_team, home_score, away_score, attendance, commissioner_cup (boolean) wnba_player_availability: avail_id, game_id, player_id, status (active/out/doubtful/questionable), reason, detected_at, official_report_time, star_tier wnba_player_stats: stat_id, game_id, player_id, minutes, points, rebounds, assists, steals, blocks, turnovers, fg_pct, 3p_pct, ft_pct, plus_minus, usage_rate wnba_team_stats: stat_id, game_id, team_id, pace, off_rating, def_rating, net_rating, efg_pct, tov_pct, orb_pct, ft_rate wnba_schedule: game_id, date, time, home_team, away_team, rest_days_home, rest_days_away, travel_distance, timezone_change, back_to_back (boolean) wnba_injuries: injury_id, player_id, injury_type, status, first_reported, last_updated, games_missed, severity_estimate wnba_overseas: overseas_id, player_id, league, team, season, games, ppg, rpg, apg, notes wnba_draft: draft_id, year, pick, player_name, college, projected_pick (consensus), workout_reports (JSON) wnba_awards: award_id, season, award_type (MVP/ROY/DPOY/6W), player_id, votes, rank, narrative_score wnba_odds: odds_id, game_id, market_type, book, selection, odds, timestamp wnba_referee: ref_id, game_id, referee_name, home_cover_pct, over_pct, foul_rate

Phase 2: Custom Metrics

Phase 3: 7 Edge Scanners

Phase 4: Matchup Card

GAME: [Away] @ [Home] | [Date] [Time ET]
ODDS: ML [home/away] | Spread [line] | Total [num]

HOME: [Team] ([Record]) | Coach: [Name] ([W-L])
  Rest: [days] | Travel: [distance] | Charter: [Y/N]
  ORtg: [val] | DRtg: [val] | Pace: [val] | Net: [val]
  Availability: [list of OUT/GTD with star tier]
  Star Impact: [rating if star absent]
  
AWAY: [Team] ([Record])
  [Same format]

CONTEXT:
  Commissioner's Cup: [Y/N]
  Expansion team: [Y/N — which]
  Ref crew: [Names] | Historical: Home cover [%], O/U [%]
  
INTELLIGENCE:
  [Overseas form notes for key players]
  [Availability detection timestamp vs market timestamp]

Phase 5: Dashboard

• Game board: upcoming games with availability status, edge flags per market
• Player availability tracker: real-time status with scratch velocity timing
• Team drill-down: on/off splits, pace, ratings, form trends
• MVP/ROY tracker: running stats + narrative + voting projections
• Draft board: mock draft consensus, workout reports, GM signals
• Referee tendencies: home cover %, over %, foul rates per crew
• Expansion team monitor: calibration progress, confidence intervals
• Edge alerts: by market type and magnitude
• P&L: by market type, edge bucket, Brier scores

Phase 6: Kill Switch

• Star player scratch <90min pre-game: Re-run model, compare to stale Kalshi price
• Expansion team early season: Reduce sizing to 50% until 15 games played
• Olympic break: Pause all season-long scanners, track who's absent
• Pinnacle not covering game: Fall back to synthetic sharp line from analyst consensus
• Correlated exposure cap: Max 2 edges from same game, adjust sizing for correlation

OPEN QUESTIONS FOR BOSS RULING

1. Pinnacle WNBA coverage: Test how many games Pinnacle actually covers. If <50%, is synthetic sharp line viable?
2. Overseas data sourcing: Which leagues to track? EuroLeague Women, Australian WNBL, Turkish league?
3. NBA priors: Use NBA model structure as starting point, or build WNBA-specific from scratch?
4. Draft scanner priority: Draft #1 pick market very thin on Kalshi. Worth building?
5. MVP narrative tracking: Automate media sentiment or manual input?
6. Paper trading period: Run 3 weeks paper before live? Entire first month?
7. Expansion team handling: Skip expansion team games or bet with reduced sizing?

COUNCIL METADATA