Golf 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: Opus (2 of 5 peer review votes — most split council, all others self-voted) Status: PENDING BOSS RULING on open questions

COUNCIL SUMMARY

Where Advisors Agreed

1. Strokes Gained is THE core framework — SG:OTT, SG:APP, SG:ARG, SG:PUTT decomposition drives everything
2. SG:Approach is the most stable/predictive component — low variance week-to-week, best single predictor
3. SG:Putting is the noisiest component — high variance, low predictability, grass type matters hugely
4. Course fit score from regression coefficients on SG components per venue
5. Weather wave advantage is the #1 short-term edge — AM vs PM wave can be 2-3 strokes
6. Monte Carlo tournament simulation (50K-100K iterations) for outright/top-N/make-cut markets
7. 7 edge scanners required — Outright, H2H, Make Cut, Top 5/10/20, Round Leader, 3-Ball, Hole in One
8. DataGolf as primary SG data source (comprehensive, covers all tours)
9. NWS API for weather (US events), separate source for international
10. Pinnacle de-vig via Power method for N-way outright markets

Where Advisors Disagreed

1. SG weighting windows: Different decay factors and lookback windows proposed. Council verdict: Recent 8 rounds (0.85 decay) 30%, Medium 24 rounds (0.92 decay) 35%, Long-term 100 rounds (0.97 decay) 35%.
2. Simulation count: Range from 50K to 100K. Council verdict: 100K for outrights (150+ player fields need it), 50K acceptable for H2H/3-ball.
3. Tour-specific modeling: Some treated all tours the same. Council verdict: Separate parameters for PGA Tour, LIV (shotgun starts → no wave advantage), DP World Tour, Korn Ferry (higher variance/sigma), LPGA (separate regressions).
4. Architecture complexity: One advisor recommended Kafka/Kubernetes/Snowflake. Council verdict: SQLite WAL, simple cron-based pipeline.

Strongest Arguments (from peer review)

Opus wins (split council — 2 genuine cross-model votes) with the most operationally specific design:

• Concrete numerical parameters throughout (decay factors, weighting splits, wind thresholds)
• Kelly fraction sizing per market type (0.25x outright, 0.5x H2H, 0.4x make-cut)
• Course DNA vector with regression-derived weights per venue (Augusta example: OTT 0.20, APP 0.30, ARG 0.30, PUTT 0.20)
• Course history adjustment caps (+0.3 max to prevent overfitting)
• Field/tee time workflow with exact ET times and polling frequencies
• WD handling for each market type
• Wind impact with 10-mph baseline threshold
• Player sigma values by archetype (2.3 elite consistent, 3.3 bombers)
• LIV shotgun starts eliminating wave advantage
• 3-analyst panel split (Fundamentals / Situational / Market)

Biggest Blind Spot

No backtesting framework or calibration methodology — All advisors describe how to generate edges but none describe how to validate whether those edges actually make money. No historical backtest specification, expected hit rates, calibration curves, or feedback loops.

What Everyone Missed (from peer reviews)

1. Hierarchical correlation structure — Players sharing the same wave, same group, same pin positions have correlated outcomes beyond weather. Need: round-level course effect → wave-level weather → group-level micro-conditions → individual noise. Models treating players as independent draws underestimate variance on wave-wide blowouts and overstate confidence on H2H/3-ball edges.
2. Pin position impact — Daily pin placements change hole difficulty by 0.5-1.0 strokes. Not captured in any SG lookback.
3. Putting surface type transitions — Bermuda → bentgrass → poa annua transitions cause huge putting variance. Players' putting SG needs grass-type-specific tracking.
4. Green firmness and course setup changes — Courses play softer after rain, harder as week progresses. Stimpmeter readings and course setup decisions affect all markets.
5. Alternate field compositions — Monday qualifiers and alternates have different SG profiles than the committed field. Pipeline must handle late additions.

BUILD PLAN

Phase 1: Core Golf Data Tables

golf_players: player_id, full_name, tour (PGA/LIV/DPWT/KF/LPGA), nationality, age, status, sg_total, sg_ott, sg_app, sg_arg, sg_putt, owgr_rank, datagolf_rank, sigma (scoring variance), updated_at

golf_player_sg: player_id, date, window (8/24/100 rounds), sg_total, sg_ott, sg_app, sg_arg, sg_putt, rounds_played, ewma_total, course_fit_score (per-tournament)

golf_courses: course_id, name, city, state_country, par, yardage, grass_type (bermuda/bentgrass/poa), altitude_ft, course_dna_ott, course_dna_app, course_dna_arg, course_dna_putt, avg_winning_score, cut_line_history_avg

golf_tournaments: tournament_id, name, course_id, tour, start_date, end_date, purse, field_size, format (stroke/match), cut_rule (top-65/top-70/no-cut), shotgun_start (boolean for LIV)

golf_field_lists: tournament_id, player_id, entry_type (committed/alternate/MQ), wd_status, wd_timestamp, tee_time_r1, wave_r1 (AM/PM), tee_time_r2, wave_r2, made_cut (boolean), final_position

golf_weather: tournament_id, round_number, wave (AM/PM), forecast_time, temperature_f, wind_speed_mph, wind_gust_mph, wind_direction, precipitation_prob, precipitation_type, wave_advantage_strokes, source (NWS/other)

golf_course_history: player_id, course_id, rounds_played, avg_sg_total, best_finish, made_cuts, missed_cuts, wins

golf_round_scores: tournament_id, round_number, player_id, score_to_par, sg_total, sg_ott, sg_app, sg_arg, sg_putt, tee_time, wave, position_after_round

Phase 2: Derived Metrics

Phase 3: Weather Model (Golf's Biggest Edge)

Data ingestion:

• NWS API hourly forecasts for US events (free)
• International events: weather API TBD
• Poll every 3 hours during tournament play
• Store: temp, wind speed/direction/gusts, precip prob, humidity

Wave advantage calculation:

1. Get hourly wind forecast for AM tee times (7:00-9:30 AM local) and PM tee times (12:00-2:30 PM local)
2. Average wind speed per wave window
3. Apply wind impact formula: strokes penalty = (avg_wind - 10) × 0.05 per 5mph, minimum 0
4. Wave advantage = PM_penalty - AM_penalty (positive = AM advantage)
5. Typical range: 0 to 3+ strokes in extreme conditions

Re-pricing triggers:

• Forecast change > 5mph wind speed differential between waves
• Rain probability change > 30% for either wave
• Weather delay/suspension announced

Phase 4: Monte Carlo Tournament Simulation

Parameters:

• 100,000 iterations per tournament
• Each iteration: simulate 72 holes (4 rounds × 18 holes) for each player
• Player round score ~ N(μ_player, σ_player²) where μ = adjusted SG:Total, σ = player sigma
• Apply wave weather adjustment per round
• Apply cut after round 2 (top-65 or tournament-specific rule)
• Track finish position per sim → derive all market probabilities

Per-sim outputs:

• Win probability (outright)
• Top 5/10/20 probability
• Make cut probability
• Each H2H pairing: who finishes higher
• Round leader probability per round
• 3-ball grouping: who wins the 3-ball

Phase 5: 7 Edge Scanners

Common engine:

1. Ingest Pinnacle odds
2. De-vig: Power method for N-way outrights, multiplicative for 2-way markets
3. Run Monte Carlo simulation
4. Compare model probabilities to Kalshi prices
5. Min edge: varies by market (see below)
6. Output: {tournament_id, market, selection, model_prob, kalshi_price, edge, confidence, wave_flag}

Phase 6: Matchup Card Format

TOURNAMENT: [Name] | [Course] | [Tour] | [Dates]
COURSE DNA: OTT [wt] | APP [wt] | ARG [wt] | PUTT [wt]
PAR: [val] | YARDAGE: [val] | GRASS: [type] | ALTITUDE: [ft]
CUT RULE: [top-65/70/no-cut] | FORMAT: [stroke/shotgun]

PLAYER: [Name] | Rank: OWGR [#] / DataGolf [#]
  SG Composite: [val] (Recent: [val] | Medium: [val] | Long: [val])
  SG Components: OTT [val] | APP [val] | ARG [val] | PUTT [val]
  Course Fit Score: [val] (z-score)
  Course History: [rounds at venue] | Best: [finish] | Cuts: [X/Y]
  Sigma: [val] (consistency tier: [elite/average/volatile])
  Form (Last 3 Events): [Event: Finish, Event: Finish, Event: Finish]

WAVE ASSIGNMENT: Round 1 [AM/PM] | Round 2 [PM/AM]
  Wave Advantage R1: [+/- strokes] | R2: [+/- strokes]

WEATHER:
  R1 AM: Wind [mph] [dir] | Temp [°F] | Rain [%]
  R1 PM: Wind [mph] [dir] | Temp [°F] | Rain [%]
  R2 AM: Wind [mph] [dir] | Temp [°F] | Rain [%]
  R2 PM: Wind [mph] [dir] | Temp [°F] | Rain [%]
  Weekend Forecast: [summary]

MODEL PROBABILITIES:
  Win: [%] | Top 5: [%] | Top 10: [%] | Top 20: [%] | Make Cut: [%]

STATUS:
  Equipment Changes: [Y/N — details]
  Caddie: [Name] | Change: [Y/N]
  Injury: [details if any]

INTELLIGENCE:
  [CRITICAL/MODERATE/CONTEXT findings]

Phase 7: Dashboard

• Tournament board: Field list with SG composites, course fit scores, wave assignments, edge counts
• Player drill-down: Full player card + all market edges + SG trends + course history
• Weather center: Wave advantage visualization, forecast timeline, re-pricing alerts
• Wave comparison: AM vs PM wave scoring projections with player splits
• Cut line tracker: Projected cut line with make-cut probabilities
• H2H board: All available H2H matchups with edges and wave differentials
• Leaderboard (live): Real-time leaderboard with weekend weather projections
• Edge alerts: Sorted by magnitude, filterable by market type
• P&L tracker: By market type, by tournament, by edge bucket, Brier scores

Phase 8: Kill Switch

• Weather suspension: Pause all scanners, void round leader edges for suspended round
• Course unplayable: Suspend all markets until course reopens
• Tournament cancellation: Void all positions
• Player WD pre-tee: Void outright + H2H for that player, re-run MC with N-1 field
• Late WD (in-tournament): Player = last place for all active markets

OPEN QUESTIONS FOR BOSS RULING

1. Tour scope: PGA Tour + LIV + DP World Tour + Korn Ferry + LPGA all at launch? Or PGA Tour first?
2. DataGolf subscription: Required for comprehensive SG data. Worth the cost?
3. Simulation count: 100K for outrights confirmed? Computationally feasible?
4. LIV shotgun starts: No wave advantage — different model needed. Worth building separate LIV logic?
5. Hole in One scanner: Very high variance, novelty market. Build now or defer?
6. Backtesting harness: Build calibration system before going live?
7. International weather: NWS only covers US events. What API for DP World Tour/international events?

COUNCIL METADATA