Synergy operates at the pitcher-season level, not per-pitch. It measures the residual CSW (called strikes + swinging strikes) after removing what individual pitch grades already predict. The leftover signal is the mix effect — how pitches interact.
1. The Mix Bonus
Two pitchers can have identical individual pitch grades yet produce very different results. The reason: their pitches either complement or duplicate each other. Arsenal Synergy captures this interaction effect.
Pitcher A — Complementary Arsenal
Pitcher B — Redundant Arsenal
🎧
Analogy: Imagine a band where every instrument plays in the same register. Technically competent, but no contrast, no dynamics. A great band has bass, midrange, and treble — instruments that are individually good and occupy different sonic space. Synergy measures the arrangement, not the musicians.
2. Velocity Separation
The speed gap between a pitcher's fastest and slowest offerings is one of the strongest synergy signals. Large velocity spreads force hitters to cover a wider timing window, making each pitch harder to square up.
Velocity Spectrum
3. Tunnel Contrast
Tunneling is the art of making different pitches look identical to the batter for as long as possible. Two pitches that share the same tunnel window — the path from release to about 20 feet from the plate — then diverge to different locations, create maximum deception.
The Tunnel Window
The longer two pitches share a tunnel, the less time the batter has to recognize the difference. Great synergy arsenals maximize the shared distance while maximizing the final separation. It's not enough to tunnel well — the pitches must also end up in very different spots.
Feature Categories
The 26 arsenal-level features capture four dimensions of pitch mix quality. Hover over each tag to see what it measures.
Mix Balance
pitch_type_countNumber of distinct pitch types thrownusage_entropyHow evenly pitches are distributed (high = balanced)max_usage_shareUsage % of the most thrown pitchsecondary_usage_sumCombined usage of non-primary pitches
Velocity Separation
velo_rangeGap between fastest and slowest pitch (mph)velo_spread_stdStandard deviation of pitch velocitiesfb_off_velo_gapSpeed difference between fastball and offspeedvelo_cluster_countNumber of distinct speed clusters
Movement Spread
hb_rangeHorizontal break range across arsenalivb_rangeInduced vertical break range across arsenalmovement_hull_areaArea of the convex hull in movement spacemax_movement_distMaximum pairwise movement distance between any two pitchesavg_movement_distAverage pairwise movement distance
Tunnel Contrast
tunnel_similarityHow similar pitches look in the first 20 feet of flightlate_divergenceHow much pitches separate in the final 20 feettunnel_ratioRatio of shared tunnel to final divergence (higher = more deceptive)release_consistencyHow tightly release points cluster across pitch types
4. Why Linear, Not XGBoost?
Stuff+, Location+, and Pitching+ all use XGBoost — a powerful tree-based model that thrives on millions of rows. So why does Arsenal Synergy use a simple RidgeCV (regularized linear regression)?
The answer is data. Synergy operates at the pitcher-season level. Each row is one pitcher's entire season, not one pitch. That means ~500 samples per year, not ~5 million.
The Bias-Variance Tradeoff
RidgeCV (What We Use)
Linear model with built-in regularization
Cross-validates to pick regularization strength
Each feature gets one coefficient — fully interpretable
Can't overfit with only 26 features and regularization
Test R² = 0.147 — stable and reliable
XGBoost (Tested, Rejected)
Powerful but data-hungry
Needs thousands of rows to learn stable splits
With 500 rows, trees memorize individual pitchers
Train R² looks great (~0.4+), test R² collapses
Predictions don't generalize year-over-year
📈
The math is simple: With 5 million pitches, XGBoost can find subtle nonlinear interactions in the data. With 500 pitcher-seasons, even moderate complexity leads to overfitting. RidgeCV's regularization shrinks unimportant coefficients toward zero, keeping only the features that consistently matter across years.
5. What Synergy Explains
Synergy captures the gap between how a pitcher should perform (based on individual pitch grades) and how they actually perform. Here are the two archetypes.
The Overperformer
This pitcher throws a diverse arsenal with excellent tunneling. Each individual pitch is average, but hitters struggle because they can never lock into one timing or movement pattern. The whole is greater than the sum of its parts.
The Underperformer
Each pitch individually has nasty stuff. But the cutter and fastball look the same, and the slider and sweeper overlap in movement. Hitters face fewer "different looks" than the pitch count suggests. Individually great, collectively redundant.
What the R² = 0.147 Means
An R² of 0.147 means synergy explains about 15% of the CSW variance that individual grades don't. That's a modest but meaningful signal — enough to move a pitcher 2-3 points on a 100-scale. It won't turn a bad pitcher good, but it separates otherwise similar pitchers.
14.7%
of residual CSW variance explained by arsenal composition alone
6. Positive vs Negative Synergy
Synergy is reported on a centered scale where 0 means no mix effect. Positive values mean the arsenal plays up beyond what the individual grades predict. Negative values mean the arsenal plays down despite individual quality.
The Synergy Scale
NEGATIVE SYNERGY
Arsenal plays down
Pitches overlap in velocity, movement, or both. Hitters can sit on one speed/shape and cover multiple offerings. Even great individual stuff can't overcome a predictable mix.
POSITIVE SYNERGY
Arsenal plays up
Pitches create diverse looks: wide velocity spread, different movement profiles, good tunneling. Hitters can never settle into a plan. Average stuff becomes effective through deception.
Bottom Line
Arsenal Synergy answers a question that individual pitch grades can't: does this pitcher's mix make sense?
It won't add 10 points to anyone's grade. But when two pitchers have similar Stuff+ and Location+, synergy is often what separates the one who gets results from the one who puzzles scouts. It's the arrangement of the orchestra, not the skill of the musicians.