MLB Player Props 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 (2 of 5 peer review votes; all others got 1 each — most split council yet) Status: PENDING BOSS RULING on open questions

COUNCIL SUMMARY

Where Advisors Agreed

1. 8 separate edge scanners required — Pitcher Strikeouts, Pitcher Outs, Batter Hits, Batter HRs, Batter RBIs, Batter Total Bases, Batter Runs, Stolen Bases
2. Pitcher strikeouts is the highest-EV prop — most liquid, most modelable, strongest signal-to-noise ratio
3. Negative Binomial for strikeouts (overdispersion vs Poisson), Zero-Inflated Poisson for HRs and outs
4. SP identity drives ALL batter props — opposing pitcher quality is the primary input for every batter prop
5. FanDuel alt lines as foundation for building probability curves (same as NHL player props)
6. Platoon splits are critical — L/R matchups create huge wOBA swings for hitter props
7. Statcast data (exit velocity, barrel rate, launch angle) essential for HR and hits modeling
8. Lineup position matters — batting order affects plate appearances, RBI/Runs opportunities
9. Separate pitcher matchup card from batter matchup card
10. 3-pass research schedule aligned with MLB game desk

Where Advisors Disagreed

1. Total Bases distribution: Grok proposed multinomial (1B/2B/3B/HR outcomes), gpt-oss proposed custom continuous, Gemini proposed linear formula (wrong — can't price derivatives). Council verdict: Multinomial is correct — each at-bat produces discrete base outcomes.
2. RBI/Runs modeling approach: Gemini used linear multipliers (broken — no distribution curve). Others used conditional models on lineup context. Council verdict: Monte Carlo game simulation (10K runs) is required — RBI and Runs are sequence-dependent, not isolated.
3. Outs recorded distribution: Sonnet proposed truncated normal from empirical game logs. Grok proposed zero-inflated truncated Poisson. Council verdict: Truncated normal is pragmatic for outs (QS pitchers cluster at 18+, bad starts at 9-12).
4. Phased rollout: Opus recommended starting with K props only and expanding. Others proposed all 8 simultaneously. Council verdict: Phase — start with pitcher Ks and outs (most modelable), then batter hits/HRs, then RBI/Runs/TB/SB.

Strongest Arguments (from peer review)

Grok wins with the sharpest prop-specific distribution choices and the most production-ready scanner designs:

• Correct distribution assignment for every prop type (NB for Ks, ZIP for HRs, multinomial for TB)
• Clean database schema with calibration table for backtesting
• Explicit automated refresh triggers per prop type
• Stays focused on the pipeline question without bloat

Opus runner-up with deepest operational insights:

• Phased rollout recommendation (K props first)
• Prop quality hierarchy by modelability and signal-to-noise
• Specific exploitable asymmetries (catcher pop time for SB, umpire K-zone, weather repricing lag)
• Only advisor using SQLite (matching actual tech stack)

Biggest Blind Spot

Gemini: Used linear arithmetic formulas instead of probability distributions for prop pricing. You cannot price an "Over 1.5 Total Bases" Kalshi contract with a mean estimate — you need a full distribution curve. This is a fundamental misunderstanding of derivative pricing.

What Everyone Missed (from peer reviews)

1. Monte Carlo for sequence-dependent props — RBI and Runs require full play-by-play game simulation, not isolated player models. A batter's RBI depends on the OBP of hitters batting before them.
2. Portfolio-level correlation within a game — Multiple props in the same game are driven by the same game script. Need intra-game covariance estimation and exposure caps per game.
3. Opener/bulk reliever dynamics — Modern MLB uses "openers" (1-inning relievers followed by bulk pitchers). Pipeline must detect these and not build cards around the 1-inning opener.
4. Platoon pinch-hit risk — Lefty batters may get pinch-hit in the 7th inning against a lefty reliever, reducing projected plate appearances and invalidating over projections.
5. Model calibration loop — Rolling-window validation, model drift detection, prop-type-specific Brier scores.

BUILD PLAN

Phase 1: Pitcher Props Data (START HERE)

mlb_pitcher_prop_logs:

• pitcher_id, date, game_id, opponent
• strikeouts, outs_recorded, innings_pitched, pitches_thrown
• k_per_9, k_rate, swinging_strike_rate, called_strike_rate
• pitch_mix (FB%, SL%, CH%, CB%, CUT% — JSON)
• first_pitch_strike_rate, chase_rate_induced
• Source: MLB Stats API + Baseball Savant

mlb_pitcher_prop_baselines:

• pitcher_id, date, stat_type (strikeouts/outs)
• last_3, last_5, last_10, season_avg
• ewma_015 (fast), ewma_010 (standard)
• vs_lhb_k_rate, vs_rhb_k_rate (platoon K rates)
• home_k_rate, away_k_rate
• early_season_flag

mlb_pitcher_matchup_context:

• pitcher_id, game_id, opponent_team
• opp_team_k_rate (how much opponent strikes out)
• opp_team_chase_rate, opp_team_contact_rate
• opp_lineup_wrc_plus, opp_lineup_k_rate_vs_hand
• umpire_k_zone_adjustment
• park_k_factor

Phase 2: Batter Props Data

mlb_batter_prop_logs:

• batter_id, date, game_id, opponent_sp_id
• hits, home_runs, rbis, runs, total_bases, stolen_bases
• plate_appearances, at_bats, walks
• exit_velocity_avg, barrel_rate, launch_angle_avg, hard_hit_rate
• vs_sp_hand (L/R), batting_order_position

mlb_batter_prop_baselines:

• batter_id, date, stat_type
• last_5, last_10, last_20, season_avg
• ewma_012, ewma_008
• vs_lhp_rate, vs_rhp_rate (platoon splits per stat)
• home_rate, away_rate
• woba, xwoba, iso_power

mlb_batter_matchup_context:

• batter_id, game_id, opponent_sp_id
• sp_pitch_mix_vs_hand (JSON — how SP attacks this batter handedness)
• sp_k_rate_vs_hand, sp_whip_vs_hand
• park_hr_factor_for_hand (L/R specific)
• weather_hr_adjustment
• lineup_position, projected_plate_appearances

Phase 3: Distribution Models Per Prop

Phase 4: Edge Scanners (8 scanners)

Common engine:

1. Ingest FanDuel alt lines (3+ thresholds per prop)
2. De-vig each threshold (multiplicative 2-way)
3. Fit distribution curve to de-vigged probabilities
4. Compare fitted curve to Kalshi contract prices
5. Calculate edge after 7% Kalshi fee
6. Min edge: 4 cents, min sample: 10 games (pitcher), 20 games (batter)
7. Output: {player, prop_type, threshold, model_prob, kalshi_price, edge, confidence}

Per-prop scanner unique logic:

Phase 5: Matchup Cards

Pitcher Prop Card:

PITCHER: [Name] ([L/R]) | TEAM: [vs OPP] | Status: [Confirmed/Probable]
RECENT (Last 5 Starts): K/9: [avg] | IP/Start: [avg] | Pitches/Start: [avg]
SEASON: K Rate: [%] | SwStr%: [%] | CSW%: [%] | WHIP: [val]
PITCH MIX: FB [%] | SL [%] | CH [%] | CB [%]
vs LHB: K Rate [%] | wOBA [val] | vs RHB: K Rate [%] | wOBA [val]

OPP LINEUP:
  Team K Rate: [%] | Chase Rate: [%] | Contact Rate: [%]
  Lineup K Rate vs [L/R]HP: [%]
  Key Hitters (top 3 K-resistant): [names + K rates]

UMPIRE: [Name] | K Zone: [+/- above avg] | ABS Overturn: [rate]
PARK: K Factor: [val]

INTELLIGENCE: [CRITICAL/MODERATE/CONTEXT findings]

Batter Prop Card:

BATTER: [Name] | POS: [pos] | BATS: [L/R/S] | TEAM: [vs OPP]
LINEUP: [Order position] | Confirmed: [Y/N]

RECENT (Last 10 GP):
  H/G: [avg] | HR/G: [avg] | RBI/G: [avg] | TB/G: [avg] | SB: [total]
  Exit Velo: [avg] | Barrel Rate: [%] | Hard Hit: [%]

SEASON RATES:
  wOBA: [val] | xwOBA: [val] | ISO: [val] | K Rate: [%] | BB Rate: [%]
  vs LHP: wOBA [val] | ISO [val] | vs RHP: wOBA [val] | ISO [val]

MATCHUP (vs [SP Name] [L/R]):
  SP K Rate vs [L/R]HB: [%] | SP WHIP vs [L/R]HB: [val]
  SP Pitch Mix vs [L/R]HB: FB [%] SL [%] CH [%]
  Batter perf vs SP pitch types: [weakness flags]

CONTEXT:
  Park HR Factor ([L/R]HB): [val] | Weather HR Adj: [+/-]
  Projected PA: [val] | Lineup Protection: [next batter wRC+]
  Platoon PH Risk: [Y/N — if lefty batter, lefty reliever likely]

INTELLIGENCE: [findings]

Phase 6: Dashboard

• Prop board: All today's player props with edge counts by prop type
• Pitcher K focus: All SPs with K prop edges, K rate trends, umpire K zones
• Batter drill-down: Click player → batter card + all prop edges + Statcast charts
• Edge alerts: Sorted by magnitude, filterable by prop type
• Lineup status: Confirmed vs projected lineups with delta alerts
• P&L tracker: Performance by prop type, by edge bucket, Brier scores
• Calibration dashboard: Rolling accuracy by prop type, model drift alerts

OPEN QUESTIONS FOR BOSS RULING

1. Phased rollout: Council recommends starting with pitcher Ks and outs (most modelable), then batter hits/HRs, then RBI/Runs/TB/SB. Confirm?

2. Monte Carlo for RBI/Runs: These props require full game simulation (10K runs) because they're sequence-dependent. This is computationally expensive. Build now or defer?

3. Stolen bases prop: Very low base rate, high variance, thin Kalshi liquidity. Worth building a scanner for, or skip initially?

4. Opener detection: Should the system auto-flag games where team is using an opener (1-2 inning reliever) vs traditional starter? This changes ALL prop projections.

5. Platoon pinch-hit risk: Should we model the probability that a lefty batter gets pinch-hit in late innings against lefty relievers? This caps plate appearances and affects all over projections.

6. Cross-desk data sharing: Pitcher prop desk should consume game desk's SP quality metrics and team batting data rather than building its own. Confirm shared data model?

COUNCIL METADATA