mlb-matchup-analyzer（MLB 比赛分析器）

MLB Matchup Analyzer

Table of Contents

• Example
• Workflow
• Signal Outputs
• Guardrails
• Quick Reference

Example

Scenario: User needs to decide whether to start Junior Caminero (TB, RHB) for today's game. Game: Tampa Bay Rays @ Colorado Rockies, 2026-04-17, Coors Field.

Research pass (web search, cite every URL):

• MLB.com probable pitchers -> COL SP is German Marquez (RHP), 2026 ERA 5.10, K% 19.0%, hard-hit% 43%
• FanGraphs park factors 2026 -> Coors Field: wOBAcon 1.18 (league-leading hitter park)
• RotoWire weather (Denver, 7:10 PT first pitch) -> 82F, 0% precip, wind out to CF 9 mph
• RotoBaller closer chart -> COL bullpen: closer healthy, setup mix stable; TB bullpen: closer used 3 of last 4, likely unavailable

Normalization pass:

• opp_sp_quality (COL SP viewed from TB hitters' side): Marquez is below-average -> low hitter-opposition. Score = 30 (50 = league-average starter; lower = easier matchup for hitters).
• park_hitter_factor: Coors wOBAcon 1.18 -> normalize so 1.00 = 50. Score = 74 (very hitter-friendly).
• park_pitcher_factor: inverse of above. Score = 26 (very unfriendly to pitchers).
• weather_risk: 0% precip + mild temp + supportive wind -> 8 (very low disruption risk).
• bullpen_state (for the home team, COL): closer rested, setup healthy -> 65. (For TB: closer gassed -> 35.)

Platoon narrative (plain English, no jargon): "Caminero hits right-handed and faces a right-handed pitcher. This is a neutral handedness matchup, not a platoon advantage. However, Caminero's hard-hit rate vs RHP is above his season average, and Coors Field is the best hitter park in baseball, so the overall matchup is strongly positive."

Composed downstream: mlb-player-analyzer reads these signals and produces matchup_score = 40% opp_sp + 25% park_hitter + 25% platoon + 10% weather -> ~72 for Caminero today. That feeds daily_quality -> START.

Workflow

Copy this checklist and track progress:

MLB Matchup Analysis Progress:
- [ ] Step 1: Collect game identifiers (home, away, date, first pitch time)
- [ ] Step 2: Identify both probable starting pitchers + handedness
- [ ] Step 3: Pull park factors (home park)
- [ ] Step 4: Pull weather forecast for first-pitch window
- [ ] Step 5: Assess bullpen state for both teams
- [ ] Step 6: Normalize each signal to 0-100 (50 = neutral)
- [ ] Step 7: Write platoon narrative (plain English)
- [ ] Step 8: Emit signal file and validate

Step 1: Collect game identifiers

• Home team, away team, local game date, first-pitch time, venue
• Confirm the game is not already postponed or rescheduled

Step 2: Identify probable starters

Pull from MLB.com probable pitchers (https://www.mlb.com/probable-pitchers). See resources/methodology.md for search procedure and fallbacks.

• Home SP: name, handedness (L/R), season ERA, K%, xFIP
• Away SP: same fields
• Role certainty: is the start confirmed or only projected?

Step 3: Pull park factors

Use FanGraphs 2026 park factors (guts.aspx?type=pf). See resources/methodology.md for the normalization formula (raw factor -> 0-100 scale).

• Composite park factor (hitter perspective)
• Composite park factor (pitcher perspective)
• Handedness-specific factors (L/R) if material

Step 4: Pull weather forecast

Use RotoWire weather (rotowire.com/baseball/weather-forecast.php). See resources/methodology.md for the risk formula.

• Precipitation probability at first-pitch window
• Temperature + wind direction + wind speed
• Is this a dome / retractable-roof park? (overrides weather)
• Game importance multiplier (standings-critical games face more delay/reschedule friction)

Step 5: Assess bullpen state

Use RotoBaller closer depth chart + last-7-day usage reports. See resources/methodology.md.

• Closer availability (appearances in last 3 days, pitch counts)
• High-leverage setup arms available
• Any IL additions or demotions in last 72 hours

Step 6: Normalize signals

Every signal must land on a 0-100 scale where 50 = league-average / neutral. See resources/methodology.md for each signal's formula.

• opp_sp_quality computed from each side's perspective
• park_hitter_factor anchored so neutral park = 50
• park_pitcher_factor = 100 - park_hitter_factor (approximate inverse)
• weather_risk = rain_prob_pct x importance_multiplier, capped at 100
• bullpen_state computed per team (home + away)

Step 7: Write platoon narrative

Per CLAUDE.md rule 5: jargon-free or translated inline. For each key hitter of interest, state handedness and the SP's handedness, then describe the platoon edge or lack thereof in plain English. See resources/template.md.

Step 8: Emit and validate

Write to signals/YYYY-MM-DD-matchup.md using resources/template.md. Call mlb-signal-emitter for validation. Validate against resources/evaluators/rubric_mlb_matchup_analyzer.json. Minimum standard: average score 3.5+.

Signal Outputs

Plus a narrative block covering platoon implications for both lineups.

Guardrails

1. 50 is neutral, always. Every signal is anchored so 50 = league-average. If your formula produces a distribution that doesn't hit 50 at the median, it's wrong -- re-anchor.
2. Cite every URL. Every signal value must be traceable to a specific source URL in source_urls:. If a fact cannot be verified via web search, drop confidence to 0.3 or lower and flag in red team.
3. Park factors are context-dependent. Handedness splits matter: Yankee Stadium favors LHB; Fenway's Green Monster favors RHB pull-hitters. If a key hitter has extreme splits, use the handedness-specific park factor, not the composite.
4. Weather = rain probability x importance multiplier. A 30% rain chance matters more in a September pennant race than in April. See methodology for the importance multiplier.
5. Bullpen state is per team, not per game. Emit two values: bullpen_state_home and bullpen_state_away. Downstream consumers pick the one they need (e.g., the streaming-strategist cares about the opposing bullpen for a late-game lead).
6. Platoon narrative must be jargon-free. Never write "positive splits vs RHP." Write "hits right-handed pitchers better than left-handed pitchers." Translate every stat the first time it appears.
7. Dome overrides weather. If the home park is a dome (or the roof is confirmed closed), set weather_risk = 0 regardless of forecast, and note the roof status in the signal body.
8. Both SPs, both sides. The signal file should report opp_sp_quality from both perspectives: the home team's hitters face the away SP, and vice versa. Do not pick only one.

Quick Reference

Key formulas:

park_hitter_factor = 50 + (raw_wOBAcon_factor - 1.00) * 200
  (clamped to [0, 100]; neutral park wOBAcon ~ 1.00 -> 50)

park_pitcher_factor = 100 - park_hitter_factor (first-order approximation)

opp_sp_quality = 50 + (lg_avg_xFIP - SP_xFIP) * 15
  (better SP = higher score = tougher on opposing hitters)

weather_risk = min(100, rain_prob_pct * importance_multiplier)
  importance_multiplier: 1.0 (early season) .. 1.5 (playoff push) .. 2.0 (postseason)

bullpen_state = 50
  - 15 if closer unavailable
  - 10 per high-leverage arm unavailable
  + 5 if closer had 2+ days rest and full bullpen rested
  (clamped [0, 100])

Key sources (see context/frameworks/data-sources.md for the full list):

• Probable pitchers: https://www.mlb.com/probable-pitchers
• Park factors: https://www.fangraphs.com/guts.aspx?type=pf&season=2026
• Weather: https://www.rotowire.com/baseball/weather-forecast.php
• Bullpen: https://www.rotoballer.com/mlb-saves-closers-depth-charts/226767
• Platoon splits: FanGraphs player page -> Splits tab

Inputs required: home team, away team, date (YYYY-MM-DD), first-pitch time (optional).

Outputs produced: one signal file at signals/YYYY-MM-DD-matchup.md with frontmatter + matchup summary table + platoon narrative + source URLs.

Key resources:

• resources/template.md: Signal file template with YAML frontmatter, matchup summary table, platoon narrative block
• resources/methodology.md: Where to search, normalization formulas, park factor conversion, weather risk computation, bullpen assessment
• resources/evaluators/rubric_mlb_matchup_analyzer.json: 8-criterion quality rubric