mlb-trade-evaluator

MLB Trade Evaluator

Table of Contents

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

Example

Scenario: Opponent offers Aaron Judge (OF) for Bobby Witt Jr. (SS) + Spencer Strider (SP). Today is 2026-06-10. User is 4th in 12-team league. cat_pressure signals from mlb-category-state-analyzer: SB pressure 85, HR pressure 55, K pressure 70, QS pressure 80, OBP pressure 60; others 40-50. User has spare OF depth but thin SS. Opponent archetype from mlb-opponent-profiler: active.

Rest-of-season ATC projections (weeks 11-23, ~105 games remaining):

Player	R	HR	RBI	SB	OBP	K	QS	SV	$ value
Judge (incoming)	75	28	78	2	.405	—	—	—	$34
Witt (outgoing)	70	20	65	25	.355	—	—	—	$32
Strider (outgoing)	—	—	—	—	—	165	12	0	$26

Per-category delta (ours minus theirs, weighted by cat_pressure):

Cat	Delta (raw)	Pressure	Weighted Δ
R	+5	45	+2.3
HR	+8	55	+4.4
RBI	+13	45	+5.9
SB	−25	85	−21.3
OBP	+.050 (×PA)	60	+3.0
K	−165	70	−115.5
QS	−12	80	−96.0
SV	0	45	0
ERA/WHIP	neutral	50	0
Sum			−217.2

Raw trade value delta = $34 − ($32 + $26) = −$24.

Slot-value bonus (principle #9): 2 players out, 1 player in → we clear +1 net slot; they clear 0. slot_value_delta = (1 − 0) × $2.50 = +$2.50.

Adverse-selection pipeline (delegated to @skills/adverse-selection-prior/):

• Inputs: offer_type=trade, proposer_archetype=active, offer_symmetry_score=20 (pre-adj ratio −37%, clearly lopsided against us), proposer_info_asymmetry=55.
• Output from skill: prior_ev_probability=0.30, recommended_adjustment=0.85, rationale cites active trader selecting this offer from many possibilities despite surface lopsidedness against us.

Adjusted trade value delta:

trade_value_delta_raw       = −$24.00
+ slot_value_delta          = +$2.50
= trade_value_delta_pre_adj = −$21.50
× recommended_adjustment    = × 0.85
= trade_value_delta_adjusted= −$18.28

Express as % of outgoing dollars: −$18.28 / $58 = −31.5% → below the −20% REJECT threshold.

Positional flex delta: −30 (lose a scarce SS, gain a surplus OF). Playoff impact (weeks 21-23): Judge's schedule has 19 games vs Witt's 18 + Strider's 4 starts → −35 (lose 4 starts of Ks/QS during championship weeks).

Verdict ladder (principle #8):

• delta_pct = −31.5%, clearly below −20%.
• Adverse-selection prior is 0.34 (strong evidence offer is −EV for us).
• Advocate and critic agree: SB, K, QS all crushed at high pressure.
• Triggers the narrow REJECT case: delta < −20% AND clear adverse-selection evidence.

Verdict: REJECT. Even the always-counter rule yields: we include the counter we would have sent — "Judge + Webb + Ruiz for Witt + Strider" — as the walk-away ask, but pre-commit to reject because the haircut keeps it underwater.

Workflow

Copy this checklist and track progress:

Trade Evaluation Progress:
- [ ] Step 1: Parse the offer (players in, players out, both sides)
- [ ] Step 2: Pull rest-of-season ATC projections for every player
- [ ] Step 3: Compute per-category deltas (counting + ratio)
- [ ] Step 4: Apply cat_pressure weights from category-state-analyzer
- [ ] Step 5: Compute raw trade_value_delta in dollars
- [ ] Step 6: Compute slot_value_delta (optionality bonus)
- [ ] Step 7: Invoke @skills/adverse-selection-prior/ and apply the haircut
- [ ] Step 8: Assess positional_flex_delta
- [ ] Step 9: Compute playoff_impact (weeks 21-23 only, July+)
- [ ] Step 10: Render verdict + counter via the +15% / -20% ladder

Step 1: Parse the offer

Record exactly who is being given up and who is being received, on both sides. Trades are zero-sum within the league. See resources/template.md for the input schema.

• List all players moving to our roster (IN)
• List all players moving off our roster (OUT)
• Note each player's Yahoo position eligibility (C/1B/2B/3B/SS/OF/Util/SP/RP)
• Note any IL status, trade protection clauses, or two-start weeks in flight
• Capture the opponent's team slug so we can read context/opponents/<team>.md for their archetype

Step 2: Pull rest-of-season projections

The authoritative source is FanGraphs ATC (ensemble projection). Fallback: FanGraphs Depth Charts, then Razzball Player Rater. For every player involved, get projected totals for remaining season.

• For hitters: rest-of-season R, HR, RBI, SB, OBP, and PAs
• For pitchers: rest-of-season K, IP, ERA, WHIP, QS, SV
• Cite each source URL in the signal file frontmatter
• If a projection is missing, drop confidence to 0.5 and flag

See resources/methodology.md#projection-sourcing for detailed source order.

Step 3: Compute per-category deltas

For counting stats (R, HR, RBI, SB, K, QS, SV): simple sum of incoming minus outgoing.

For ratio stats (OBP, ERA, WHIP): weight by projected PAs (batters) or IP (pitchers). See resources/methodology.md#ratio-category-math for the weighting formula.

• Counting stat deltas per cat
• Ratio cat deltas expressed as a shift in team average, weighted by volume
• Record each in a per-cat delta table (both teams, side by side)

Step 4: Apply cat_pressure weights

Read the most recent signals/YYYY-MM-DD-cat-state.md from mlb-category-state-analyzer. For each of the 10 cats, multiply the raw delta by cat_pressure / 50.

• Weighted delta = raw_delta × (cat_pressure / 50)
• Sum the weighted deltas → cat_pressure_weighted_delta

Step 5: Compute raw trade_value_delta

Use FanGraphs Auction Calculator rest-of-season dollar values (or Razzball as fallback).

trade_value_delta_raw = Sum($ of players IN) − Sum($ of players OUT)

This is the starting point — it does NOT yet include the slot-value bonus or the adverse-selection haircut.

Step 6: Compute slot_value_delta (optionality bonus)

Every open bench slot is worth ~$2-3 in optionality: it can host a streamer, an IL stash, or a handcuff speculation. A 2-for-1 trade frees a roster slot on one side.

slot_value_delta = (N_slots_cleared_for_us − N_slots_cleared_for_them) × $2.50

Where:

• N_slots_cleared_for_us = (players OUT from our side) − (players IN to our side), clamped at ≥ 0
• N_slots_cleared_for_them = mirror on their side

Examples:

• 1-for-1: 0 slots cleared either side → slot_value_delta = $0
• 2-for-1 (we send 2, get 1): +1 slot us, 0 them → +$2.50
• 1-for-2 (we send 1, get 2): 0 us, +1 them → −$2.50
• 3-for-1 (we send 3, get 1): +2 us, 0 them → +$5.00

Add to trade_value_delta_raw to form trade_value_delta_pre_adj.

See resources/methodology.md#slot-value-optionality for the rationale (game-theory-principles.md #9).

Step 7: Invoke @skills/adverse-selection-prior/ and apply the haircut

This step operationalizes principle #4 (adverse selection on incoming offers). The evaluator does NOT compute the prior itself — it delegates to the dedicated skill and consumes the output.

• Prepare inputs for @skills/adverse-selection-prior/:
  • offer_type: trade (or counter_offer if this is a counter to a prior proposal)
  • proposer_archetype: read from context/opponents/<opponent-team>.md (one of active, expert, dormant, frustrated, unknown). If file missing, use unknown.
  • offer_symmetry_score: integer 0-100 derived from our own surface valuation. Map via:
    • trade_value_delta_pre_adj / Σ($_OUT) ≥ 0 → 72 (offer looks fair-to-favorable)
    • 0 > ratio ≥ −10% → 60
    • −10% > ratio ≥ −25% → 40
    • ratio < −25% → 20
  • proposer_info_asymmetry: integer 0-100. Default 50. Raise toward 80+ if any of: recent injury news we have not yet verified, closer role shift pending, lineup construction change, trade-deadline timing. See resources/methodology.md#adverse-selection-delegation for the scoring guide.
• Invoke @skills/adverse-selection-prior/ with those four inputs.
• Consume the skill's output contract:
  • prior_ev_probability — store verbatim in the signal file.
  • recommended_adjustment (float, typically 0.80-1.00) — this is the multiplicative haircut.
  • bayesian_rationale — embed in the verdict block's rationale.
  • override_hints — add each hint as a red_team_finding in the signal file.
• Compute:

  trade_value_delta_adjusted = trade_value_delta_pre_adj × recommended_adjustment
  

• Record adverse_selection_adjustment in the output signal block = 1 − recommended_adjustment (expressed as a percentage haircut, e.g., 0.88 → "12% haircut").

Sign note: The haircut is multiplicative on the delta. It shrinks the MAGNITUDE of our estimate in both directions — a positive delta moves toward zero (we over-estimated the gain), and a negative delta also moves toward zero (we over-estimated the loss). This is mathematically correct for an Akerlof-style prior: the skill tells us our model is less reliable than we thought, so we should anchor more toward zero. The verdict ladder compensates by setting tight thresholds (+15% / −20%) on the post-haircut percentage, so a moderately negative raw delta with a deep haircut will land in the middle band (COUNTER), while a very negative raw delta with any haircut will still clear the −20% REJECT threshold. If the haircut creates a tie at a ladder boundary, resolve toward COUNTER (per the always-counter rule).

Step 8: Assess positional_flex_delta

Score −100 to +100 based on how the trade changes roster flexibility. See resources/methodology.md#positional-flex-scoring.

• Gaining a scarce position (C, SS, 2B) from surplus = positive
• Losing a scarce position = negative
• Gaining multi-position eligibility = positive
• Consolidating two bench players into one starter = positive (frees bench)

Step 9: Compute playoff_impact (July 1+ only)

For trades made July 1 or later, evaluate specifically for championship weeks 21, 22, 23. Read the mlb-playoff-scheduler signal file for playoff_games and playoff_matchup_quality per player.

playoff_impact = Sum(playoff_games × matchup_quality of IN)
               − Sum(playoff_games × matchup_quality of OUT)

Normalize to 0-100 where 50 = neutral.

Step 10: Render verdict via the +15% / −20% ladder

Express trade_value_delta_adjusted as a percent of outgoing dollars:

delta_pct = trade_value_delta_adjusted / Σ($_OUT)

Apply the principle #8 verdict ladder:

delta_pct	Verdict	Conditions
≥ +15%	ACCEPT	AND advocate/critic variants agree AND no cat with pressure ≥80 has negative weighted delta
−20% < delta_pct < +15%	COUNTER (with specific package)	ALWAYS produce a specific counter; never pure-reject in this band
≤ −20%	REJECT	AND prior_ev_probability ≤ 0.35 (clear adverse-selection evidence)

Always-counter rule: in the middle band, even if our instinct is to decline, we ship a specific counter-package per principle #8 (repeated-game reputation). Rejection dismissively dries up future offers; a reasoned counter keeps the pipeline open.

If REJECT conditions are only partially met (e.g., delta_pct ≤ −20% but prior_ev_probability > 0.35): downgrade to COUNTER with a more demanding package.

Every verdict ends with a single-verb recommendation: ACCEPT, COUNTER (with specific proposal), or REJECT. No "consider."

See resources/template.md#verdict-block for output schema. Validate using resources/evaluators/rubric_mlb_trade_evaluator.json. Minimum average score to ship: 3.5.

Common Patterns

Pattern 1: Star-for-depth offer

• Opponent offers one star for two or three of our mid-tier players. Opponent gains a headline name; we gain slot-value bonus because the 2-for-1 frees a bench slot.
• Typical trap: dollar values look close, but we should still expect a haircut from adverse-selection-prior because they chose this offer.
• Rule: slot_value_delta tilts toward us (+$2.50 or +$5); but the adverse-selection haircut applies — check if trade_value_delta_adjusted clears +15% AND the players we give up aren't droppable-tier. If in doubt, COUNTER (not REJECT).

Pattern 2: Category-targeted swap

• Symmetric-value trade where both sides target their weaknesses (e.g., opponent gives us K/QS for our HR/RBI).
• Diagnostic: does the swap match OUR cat_pressure, or does it match THEIRS? If theirs, haircut is deep (symmetry actually bad for us).
• Rule: only ACCEPT if the categories we gain have pressure ≥70 AND the categories we give up have pressure ≤40 AND delta_pct ≥ +15% after haircut. Otherwise COUNTER.

Pattern 3: Buy-low / sell-high regression play

• Offer involves a player currently over/underperforming their underlying metrics. Cross-reference mlb-regression-flagger signal.
• Rule: if their player's regression_index is sharply positive (unlucky, bounce-back due) AND ours is sharply negative (lucky, regression incoming), pass this to @skills/adverse-selection-prior/ as proposer_info_asymmetry of 30-40 (we plausibly have better regression info than they do). The resulting shallow haircut may let delta_pct clear +15%.

Pattern 4: Injury-adjacent desperation offer

• Opponent's player is on IL or nursing a minor injury. Offer dangles a discount.
• Rule: set proposer_info_asymmetry = 80+ (they know more about the injury). The skill will return recommended_adjustment ≈ 0.80. Usually COUNTER or REJECT depending on whether delta_pct after haircut lands below −20%.

Pattern 5: Playoff-week schedule arbitrage (July+)

• Incoming player has 20 games in weeks 21-23; outgoing has 14. Even if ROS dollar value is flat, playoff value differs.
• Rule: double-weight playoff_impact for any trade proposed July 15 or later. Use playoff_impact positive swings to tip close COUNTER cases into ACCEPT.

Guardrails

1. Always counter — pure REJECT is narrow. Per game-theory-principles.md #8 (repeated-game reputation), rejecting fairly with a counter keeps offer pipelines open; dismissive rejection dries them up. REJECT is reserved for delta_pct ≤ −20% AND clear adverse-selection evidence (prior_ev_probability ≤ 0.35). Every other middle-band offer → COUNTER with a specific package.

2. Delegate the adverse-selection prior — never recompute it inline. Step 7 invokes @skills/adverse-selection-prior/. Do not duplicate its logic here. The skill is reused across trades, waiver drops, and future M&A/negotiation skills; keeping it single-source prevents drift.

3. Apply the haircut to trade_value_delta_pre_adj, not to per-category deltas. The prior adjusts our net dollar estimate, not individual cat contributions. The cat delta table stays un-haircut and is displayed as-is in the template.

4. Slot-value bonus is symmetric. If they send two and we send one, slot_value_delta is NEGATIVE from our perspective — they got the bench-slot optionality. Do not forget the minus sign on incoming 2-for-1s.

5. Use rest-of-season projections, never full-season or season-to-date. Same rule as before; atcr not atc.

6. Ratio categories need volume weighting. OBP, ERA, WHIP deltas must be PA- or IP-weighted.

7. Read the cat_pressure signal first. If mlb-category-state-analyzer has not emitted for today, run it first.

8. Quantify the counter — don't say "ask for more." Propose a specific alternative package with named players.

9. Check for trade deadline proximity. Yahoo's deadline is August 6. Flag any trade proposed in the week before.

10. Two-way impact. If the opponent is a direct playoff-seed competitor, the trade helping them is doubly bad. Note opponent identity and their standings proximity in the output block.

11. Beginner-voice output. The user has zero baseball knowledge. Translate every stat into plain English at least once. Every user-facing recommendation ends with ACCEPT, COUNTER (with specific package), or REJECT.

12. Log the decision. Emit via mlb-decision-logger to tracker/decisions-log.md with full signal values, including prior_ev_probability, recommended_adjustment, slot_value_delta, trade_value_delta_adjusted, verdict, confidence, and will_verify_on date (4 weeks out).

Quick Reference

Key formulas:

Counting-cat delta (cat C):
  raw_delta_C = Σ(projected_C of IN) − Σ(projected_C of OUT)

Ratio-cat delta (OBP example):
  PA-weighted OBP_IN and OBP_OUT, then team-level shift

Weighted cat delta:
  weighted_delta_C = raw_delta_C × (cat_pressure_C / 50)

Total cat delta:
  trade_cat_delta = Σ weighted_delta_C

Raw trade value:
  trade_value_delta_raw = Σ($_IN) − Σ($_OUT)

Slot-value bonus:
  slot_value_delta = (slots_cleared_us − slots_cleared_them) × $2.50

Pre-adjustment delta:
  trade_value_delta_pre_adj = trade_value_delta_raw + slot_value_delta

Adverse-selection haircut (via @skills/adverse-selection-prior/):
  trade_value_delta_adjusted = trade_value_delta_pre_adj × recommended_adjustment

Percent expression:
  delta_pct = trade_value_delta_adjusted / Σ($_OUT)

Playoff impact (July+):
  playoff_impact = Σ(games_21_23 × matchup_q of IN)
                 − Σ(games_21_23 × matchup_q of OUT)

Verdict ladder (principle #8):

delta_pct	Verdict	Additional gates
≥ +15%	ACCEPT	AND advocate/critic agree AND no pressure ≥80 cat is negative
−20% < d < +15%	COUNTER (with specific package)	ALWAYS — never pure-reject in this band
≤ −20%	REJECT	AND prior_ev_probability ≤ 0.35

Adverse-selection prior summary (from @skills/adverse-selection-prior/):

recommended_adjustment	Typical cause	Effect on delta_pct
1.00	Dormant opponent, shallow info gap	No change
0.95	Active opponent, symmetric offer	−5% on magnitude
0.90	Standard active opponent	−10% on magnitude
0.85	Expert opponent, some info gap	−15% on magnitude
0.80	Expert opponent, material info asymmetry	−20% on magnitude

Key resources:

• resources/template.md: Offer input schema, per-cat delta table (both teams), value/slot/adverse-selection/flex/playoff blocks, verdict block with rationale and counter, three worked examples (ACCEPT, COUNTER, REJECT)
• resources/methodology.md: Projection sourcing, counting vs ratio cat math, cat_pressure weighting, value dollars, slot-value optionality, delegation to adverse-selection-prior, flex scoring, playoff impact, new verdict ladder with always-counter rule
• resources/evaluators/rubric_mlb_trade_evaluator.json: Quality criteria including adverse-selection integration, slot-value bonus application, and verdict-ladder correctness

Inputs required:

• Trade offer (players IN, players OUT, initiating opponent)
• Rest-of-season ATC projections (FanGraphs)
• Current cat_pressure signal (from mlb-category-state-analyzer)
• Auction-calculator dollar values (FanGraphs)
• Positional eligibility for each player (Yahoo)
• Opponent archetype from context/opponents/<team>.md (written by mlb-opponent-profiler)
• regression_index signal (from mlb-regression-flagger) for buy-low/sell-high check and info-asymmetry scoring
• Injury status (RotoWire)
• If July+: playoff_games and playoff_matchup_quality per player
• Adverse-selection output from @skills/adverse-selection-prior/ (Step 7)

Outputs produced:

• trade_cat_delta per each of 10 cats (raw + pressure-weighted)
• trade_value_delta_raw ($)
• slot_value_delta ($)
• trade_value_delta_pre_adj ($)
• prior_ev_probability (from delegated skill)
• recommended_adjustment (from delegated skill)
• adverse_selection_adjustment (= 1 − recommended_adjustment)
• trade_value_delta_adjusted ($)
• delta_pct (trade_value_delta_adjusted / Σ$_OUT)
• positional_flex_delta (±100)
• playoff_impact (0-100, July+)
• verdict (accept / counter / reject) with rationale (includes bayesian_rationale from adverse-selection skill)
• Counter package: always specified unless verdict is ACCEPT (even on REJECT, include the counter we would have sent as context)
• Signal file written to signals/YYYY-MM-DD-trade-<opponent>.md
• Decision log entry via mlb-decision-logger