mlb-player-analyzer（MLB 球员分析器）

MLB Player Analyzer

Table of Contents

• Example
• Workflow
• Common Patterns
• Guardrails
• Quick Reference

Example

Scenario: Hitter analysis for Junior Caminero (TB 3B), today's opponent BOS, opp SP Brayan Bello (RHP), park Fenway, light wind.

Inputs assembled from web search:

• Last 15 days xwOBA:.410 vs season xwOBA.355 (Baseball Savant)
• Bello 2026 K/9: 8.1, xFIP 4.05, wOBA-vs-RHH.335 (FanGraphs)
• Fenway park factor: 103 R, 105 HR (FanGraphs park factors)
• Caminero vs RHP wOBA:.360 (FanGraphs splits)
• Weather: 62F, 8mph wind LF-to-CF (RotoWire)
• Lineup: confirmed #3 hitter (MLB.com starting lineups)
• Season actual wOBA.340 vs xwOBA.370 -> unlucky +.030

Signal computation:

Recommendation to lineup-optimizer: daily_quality = 66 -> START. regression_index = +15 suggests no need to sit on any recent cold-streak noise.

Pitcher counter-example (pitcher start): Bowden Francis (TOR) at COL. daily_quality replaced by streamability_score. Coors kills streamability_score regardless of raw stuff; skill would emit qs_probability ~28, k_ceiling ~40, era_whip_risk ~82 -> streamability_score ~32 (sub-70 threshold -> SIT / DO NOT STREAM).

Workflow

Copy this checklist and track progress:

MLB Player Analysis Progress:
- [ ] Step 1: Classify player (hitter vs pitcher; SP vs RP)
- [ ] Step 2: Collect season + 15-day performance (Savant, FanGraphs)
- [ ] Step 3: Collect today's context (opp SP/hitters, park, weather, lineup)
- [ ] Step 4: Compute normalized component scores
- [ ] Step 5: Compute composite signals (daily_quality or streamability_score)
- [ ] Step 6: Check regression_index and role_certainty
- [ ] Step 7: Validate against rubric and emit signal file

Step 1: Classify player

Determine role: hitter (any position player), SP (starter), RP (reliever, closer or setup). The signal set is different per role. See resources/methodology.md for role determination rules when a player has dual eligibility (two-way player, opener + bulk).

• Confirm today's role: is the player in today's lineup? Is the SP on the probable-pitcher chart today?
• If RP: is this a save-role RP or middle-relief RP? Closer depth lookup required.

Step 2: Collect performance data

Web-search the primary sources. Every URL goes in the signal file's source_urls: list.

• Baseball Savant player page: season xwOBA, 15-day xwOBA, xBA, barrel %, hard-hit %, sprint speed (hitters); xERA, whiff %, chase %, CSW (pitchers)
• FanGraphs player page: ATC projections (rest-of-season rate stats), splits tab (vs LHP / vs RHP)
• If search fails for any metric: record the attempt, set confidence: 0.3, and note the gap in the red-team field

See resources/data-cheatsheet for exact URL patterns.

Step 3: Collect today's context

• Opp SP (from MLB.com probable pitchers or matchup-analyzer signal if already emitted)
• Park factor (from FanGraphs park factors, or matchup-analyzer's park_hitter_factor / park_pitcher_factor)
• Weather (from RotoWire weather-forecast, or matchup-analyzer's weather_risk)
• Confirmed lineup slot (from MLB.com starting-lineups; posts ~2-3h pre-game)
• If a matchup-analyzer signal file exists for today's game, consume those signals -- do not re-derive

Step 4: Compute normalized component scores

All raw stats are converted to 0-100 (unipolar) or +/-100 (bipolar) per the signal framework. See resources/methodology.md for each formula.

• Hitter: form_score, matchup_score, opportunity_score (components of daily_quality)
• Hitter: regression_index, obp_contribution, sb_opportunity, role_certainty
• Pitcher: qs_probability, k_ceiling, era_whip_risk
• Pitcher RP: save_role_certainty

Step 5: Compute composite signals

• Hitter primary: daily_quality = 0.35 * form_score + 0.40 * matchup_score + 0.25 * opportunity_score
• Pitcher SP primary: streamability_score = 0.40 * qs_probability + 0.30 * k_ceiling + 0.30 * (100 - era_whip_risk)
• Pitcher SP weekly: two_start_bonus (bool from FantasyPros two-start page)

Step 6: Check regression and role

• regression_index = clamp((xwOBA - wOBA) * 500, -100, +100). Positive = unlucky (buy). Negative = lucky (sell / fade).
• role_certainty (hitter): 100 = confirmed in today's lineup, 70 = probable per beat reporter, 40 = platoon uncertain, 0 = benched or injured
• save_role_certainty (RP): 100 = locked closer per RotoBaller, 50 = timeshare, 20 = 7th-inning guy

Step 7: Validate and emit

• Fill resources/template.md frontmatter and tables
• Score against resources/evaluators/rubric_mlb_player_analyzer.json. Target average >= 3.5
• Every numeric signal has confidence and at least one source_url
• Call mlb-signal-emitter (validation); on failure, log to tracker/decisions-log.md

Common Patterns

Pattern 1: Hot Streak Hitter (Sell-the-News)

• Profile: Rolling 15-day wOBA well above xwOBA (actual outperforming expected)
• Signal signature: form_score high (>=70), regression_index negative (e.g., -25)
• Read: Production is BABIP-aided and not backed by Statcast quality. Do not overweight recent numbers.
• Action feed to lineup-optimizer: trim daily_quality by ~5 points mentally; flag for the waiver-analyst if the user is considering selling high

Pattern 2: Cold Hitter with Loud Contact (Buy-Window)

• Profile: Rolling 15-day wOBA below season average but xwOBA still strong (>=season xwOBA)
• Signal signature: form_score depressed, regression_index positive (>=+20), barrel% still good
• Read: Bad-luck stretch. Underlying contact quality intact. Start through it.
• Action: keep daily_quality weight as computed; flag positive regression to category-strategist (this is the guy who will pop next week)

Pattern 3: Two-Start Pitcher in a Bad Park

• Profile: SP with two starts this scoring week, one of which is at COL / CIN / BOS
• Signal signature: two_start_bonus = true, but one start has era_whip_risk >= 70
• Read: Volume pays in K and QS, but a blowup in Coors could torch ERA/WHIP for the week
• Action: Emit both starts as separate pitcher signals, each with its own streamability_score; streaming-strategist decides whether to eat the bad park for the volume

Pattern 4: Closer in Committee / Role Uncertainty

• Profile: RP with save opportunities but manager has said "mix-and-match" or "matchup based"
• Signal signature: save_role_certainty <= 50, k_ceiling decent, era_whip_risk low
• Read: Rostering pays only if saves materialize. Great ratios but the fantasy cat (SV) is unreliable.
• Action: Note explicitly in the signal body. Waiver-analyst uses this to decide FAAB willingness.

Guardrails

1. Cite every fact. Every numeric input (xwOBA, projected PAs, park factor, CS%) must trace to a URL in source_urls:. Unsourced claims fail the rubric's Source Citation criterion.
2. OBP matters more than AVG for this league. Our batting cats are R/HR/RBI/SB/OBP (not AVG). When computing obp_contribution and when choosing which rate stat to weight in form_score, use OBP or wOBA (which is walk-inclusive), never AVG alone. Walk rate is a feature, not a footnote.
3. QS matters more than W for this league. For qs_probability, compute the probability of 6+ IP and <=3 ER, not the probability of a win. Ignore bullpen-game starters and openers -- they score zero QS points by definition.
4. Use ATC projections, not Steamer alone. FanGraphs ATC is the consensus ensemble and is the most accurate single source. Steamer and ZiPS can be consulted for triangulation but do not substitute ATC without noting it.
5. Degrade gracefully on search failure. If a source is unreachable, do not invent numbers. Set that component's confidence to 0.3 and record the gap in the red-team note field. The red-team pass will escalate if confidence < 0.4.
6. Do not re-derive matchup-analyzer signals. If signals/YYYY-MM-DD-matchup.md exists for today's game, consume opp_sp_quality, park_hitter_factor, park_pitcher_factor, weather_risk, bullpen_state directly. Re-deriving wastes runtime and risks inconsistency across agents.
7. Timestamp every signal. computed_at: YYYY-MM-DDTHH:MMZ. Morning-brief calls are fresh; afternoon re-checks (once lineups post) supersede the morning signal with higher role_certainty.
8. Range-check every number. 0-100 signals never exceed 100 or go negative. +/-100 signals (regression_index) are clamped. The mlb-signal-emitter validator rejects out-of-range values -- check before calling.
9. Plain-English body. The frontmatter is for machines; the body must be jargon-free or translate jargon inline for the end user. "xwOBA" -> "expected offensive output based on how hard and at what angle he hit the ball, regardless of whether balls found gloves."

Quick Reference

Composite formulas (see resources/methodology.md for derivations):

daily_quality       = 0.35 * form_score + 0.40 * matchup_score + 0.25 * opportunity_score
streamability_score = 0.40 * qs_probability + 0.30 * k_ceiling + 0.30 * (100 - era_whip_risk)
regression_index    = clamp((xwOBA - wOBA) * 500, -100, +100)

Action thresholds (feed to lineup-optimizer / streaming-strategist):

Source priority (always try in this order):

Key resources:

• resources/template.md: Signal file template with YAML frontmatter, hitter and pitcher signal tables, plain-English body, and a worked example
• resources/methodology.md: Source URL cheatsheet, per-signal normalization formulas, composite computation, regression math, confidence-assignment rules
• resources/evaluators/rubric_mlb_player_analyzer.json: 9-criterion scoring rubric

Inputs required:

• Player name (exact, with team abbreviation if ambiguous, e.g., "Will Smith (LAD)")
• Player's MLB team (3-letter abbr)
• Today's opponent SP (if known; otherwise skill will web-search MLB.com probables)
• Today's park / weather (from matchup-analyzer signal file if available)

Outputs produced:

• signals/YYYY-MM-DD-player-<lastname>-<firstinitial>.md (one file per player analyzed per day)
• Populated with all hitter or pitcher signals per signal-framework.md
• Body includes plain-English translation for the end user