Heads up: All models on this page are in their early stages and should be considered placeholders. The underlying methodology is still being refined — take the specific numbers with a grain of salt.

Data & Features

What powers the models

Every grade starts with Statcast pitch tracking data — the same system that measures velocity, spin, and movement for every MLB pitch. Features are engineered to keep command, stuff, and context cleanly separated across models.

8.9M+
Total pitches
Every MLB pitch from 2015–25
2015–25
Seasons covered
11 years of Statcast tracking data
6-way
Model split
Pitch category × batter handedness
100+
Min pitch threshold
Below this, grades are unreliable
Feature families
Physics
24 features · Used by Stuff+
  • Velocity (effective speed, plate speed)
  • Movement (horizontal break, induced vertical break, total break)
  • Spin (rate, axis, efficiency, active spin)
  • Release (extension, position, arm slot)
  • Deception (approach angle, tunnel metrics, VAA)
Location
11 features · Used by Location+
  • Plate crossing position (horizontal and vertical)
  • Zone-relative placement
  • Arm-plane coordinates (along arm, across arm)
  • Against-break location (how far the pitch fought its spin)
Count & Context
5 features · Used by Location+ & Pitching+
  • Balls and strikes in the count
  • Two-strike indicator
  • Three-ball indicator
  • Inning context (decomposed)
Sequencing
6 features · Used by Pitching+ only
  • How many times the batter has seen this pitch type
  • Tunnel rotation from the fastball
  • Axis deviation from expected spin
  • Fastball movement spread (consistency)
  • Break x location interaction
Arsenal
26 features · Used by Arsenal Synergy
  • Pitch type usage distribution
  • Velocity separation between pitches
  • Movement spread across the arsenal
  • Tunnel contrast between offerings
Data sources
Statcast + FanGraphs

Statcast provides pitch-level tracking (velocity, spin, movement, location) for every MLB pitch since 2015. FanGraphs provides season-level context (ERA, FIP, K%, BB%) for validation and projections.

Why 6-way splits?
Fastball / Breaking / Offspeed × L / R batter

A slider to a right-handed batter is a fundamentally different pitch than a slider to a lefty. Separate models let each learn cleaner surfaces without cross-contamination from different pitch physics.

Feature boundaries are intentional

Location+ sees only location and count. Stuff+ sees only physics. This isn't a limitation — it's the design. By keeping inputs separated, each model produces a clean, interpretable signal. Pitching+ then combines everything for the full picture.