Starting point

Every player starts at 1,500 on both paths. Form is floored at 1,200. Matchup/production is not floored — a floor would break zero-sum.

Form path (leaderboard ELO)

Each PA with a run-value is scored as a discrete outcome V, compared to expected E_bat, then mean-reverted:

E_bat = 1 / (1 + exp(−(bat_form − pit_form) / 400 + 0.693))
Δbat = 12 × (V − E_bat) − λ × (bat_form − 1500)
Δpit = 12 × ((1 − V) − (1 − E_bat)) − λ × (pit_form − 1500)

The 0.693 offset makes E_bat = 0.333 at equal ratings. That is expected outcome value V, not OBP. Pitchers still have the advantage. Mean-reversion λ is 0.005 for batters, 0.003 for SP, 0.007 for RP.

Matchup path (batter production)

Continuous run value, strict zero-sum. Quality of contact lives only on this path.

adjusted_rv = actual_rv − (park_factor − 1) × 0.1
rv_diff = adjusted_rv − mean_rv[base_out_state]
k_eff = EVENT_K[result] × physics_mod
Δbat = k_eff × rv_diff    Δpit = −Δbat

physics_mod = 1 + 0.3 × (xwOBA − 0.315) / 0.315, clamped to [0.70, 1.30]. Non-BIP events (K, BB, IBB, HBP, SAC, E) use 1.00. A barreled out still gets partial credit; a bloop single is discounted.

Fielding errors: if the batter would gain matchup ELO, that gain is zeroed; if the pitcher would lose, that loss is zeroed. Form still applies V = 0.45. Errors are not a free win for the batter.

Park examples

Pitcher production and skill ELO

Raw matchup pitching ELO is volume-poisoned (workhorse SPs get dragged by PA count). Public ranks never use it alone.

Production score (leaderboard Production track):
process = 1500 + (SwStr − 11) × 8 + (CSW − 28) × 1.5 clamped [1350, 1750]
mu_norm = 1500 + ((matchup − 1500) / PA) × 200 × 0.25 clamped [1300, 1700]
production = 0.55 × form + 0.35 × process + 0.10 × mu_norm

Skill ELO (matchup predictor, two-start, start/sit): form ELO plus clamp((SwStr − 11) × 3, −25, +40). No matchup residual.

Clutch (RISP) track

Parallel Form-path ratings updated only when state % 8 ≥ 2 (runner on 2nd and/or 3rd). Same α = 12 and same role λ. Shown on the Talent → Clutch board. Requires 10+ RISP PAs.

Two-way players and OHLC

Batting and pitching ratings are independent. A two-way player (30+ PA each role) appears on both boards with a TWP badge. Daily candles are Form-path Open/High/Low/Close. Green = close > open. MA5 / MA15 smooth noise.

Talent is not a third ELO path

Contact / Power / Discipline / Stuff / BIP Suppression / Command are season-median z-score blends, not plate-appearance outcome models. They replaced the retired plate-appearance outcome weights that used to drive this tab.

skill = 1500 + Σ clip(z, −2, 2) × weight + 0.20 × (form_elo − 1500), floored at 1,200. Missing metrics drop out and remaining weights are renormalized. Sample floors: 100 PA batters, 60 PA pitchers. Medians are computed from the current season pool, not hardcoded league averages.

Batter dimensions

Pitcher dimensions

What the predictor actually uses

Not raw leaderboard ELO. Not talent z-scores. The PA model was calibrated on Form-scale diffs (~±150). Raw batter production (often 2,000–2,700) would saturate the logistic, so production is compressed first.

scaled_prod = 1500 + (batter_production − 1500) × 0.15
batter_model = 0.55 × scaled_prod + 0.45 × batter_form
pitcher_model = pitcher_skill_elo (form + SwStr blend; see Engine)
diff = batter_model − pitcher_model

Output: P(K), P(BB), P(1B), P(2B), P(3B), P(HR), P(Out). Directional, not betting-calibrated.

Logistic mapping

elo_feature = diff / 100
P_raw(event) = sigmoid(intercept + scale × elo_feature + park + platoon + rest + weather)

If the non-out mass exceeds 0.98 it is scaled back to 0.98. P(Out) = 1 − Σ P_raw, then the seven events are renormalized. At diff = 0 the intercepts recover league-average rates: ~23% K, ~8.5% BB, ~3.4% HR.

Expected PA wOBA is the probability-weighted sum of those weights. Matchup Lift vs Neutral is percentage points vs diff = 0.

Context adjustments

Added to every event’s linear term. They do not change stored ELO.

• Park: (park_factor − 1) × 0.5 (Coors 1.13 → +0.065)
• Weather: (launchpad_index − park_factor) × 0.3 from the weather cache; miss is ignored
• Platoon: opposite-hand +0.10, same-hand −0.06, switch-hitter +0.06
• Rest: reserved, currently 0

Favor labels

• Single matchup: composite = elo_feature + Σ context. Batter if > 0.10, pitcher if < −0.10, else neutral. Confidence = min(|composite| / 0.5, 1).
• Pitcher vs Team: favor is relative to that batch. A batter is favored only if Pred PA wOBA is more than 0.020 above the team average in the run. Minimum 50 batting PA. Home park defaults to the opponent unless you override it.