A machine learning pipeline that predicts pitch quality from physical Statcast characteristics — replicating and extending the Stuff+ methodology used in MLB front offices.
Built as the capstone project for SABR Analytics Certification Level IV.
Can we predict the run value of a pitch based solely on its physical characteristics at the moment of release — independent of outcome?
Stuff+ is a pitch quality metric that evaluates how "nasty" a pitch is based purely on its physics: velocity, movement, spin, release point. A Stuff+ of 100 is league average. Above 100 is better. The metric deliberately ignores what happened after contact — it captures intrinsic pitch quality, not luck.
stuff-plus-model/
│
├── stuff_plus_model.ipynb # Python pipeline — full Statcast pull → model → leaderboard
├── stuff_plus_models.R # R implementation — Random Forest baseline vs. improved model
├── stuff_predictions.csv # Model output — pitch-level Stuff+ predictions
├── README.md # This file
Two implementations, one problem: The Python notebook is the primary model. The R script is a companion implementation built during the SABR coursework, showing the same modeling problem approached in R with ranger. Together they demonstrate cross-language analytical ability.
delta_run_exp — the change in run expectancy per pitch, sourced directly from MLB Statcast via pybaseball. This is a true outcome-based run value, making it an ideal training signal for pitch quality.
| Feature | Description |
|---|---|
release_speed |
Velocity at release point (mph) |
release_spin_rate |
Spin rate (rpm) |
release_extension |
Distance from rubber at release (ft) |
release_pos_x / z |
Horizontal and vertical release point |
pfx_x |
Horizontal movement vs. gravity-only path (in) |
pfx_z |
Induced vertical break / IVB (in) |
plate_x / z |
Location at the plate |
spin_axis |
Spin axis in degrees |
pitch_type |
Pitch type (encoded) |
| Feature | Rationale |
|---|---|
spin_velocity_ratio |
Spin rate per mph — captures spin efficiency, not raw spin. Pitches with high spin relative to velocity tend to have more deceptive movement. |
total_break |
√(pfx_x² + pfx_z²) — total movement magnitude combining horizontal and vertical. |
plate_dist |
Distance from strike zone center — pitches at the edges or just off the zone are harder to handle. |
velo_x_ivb |
Velocity × IVB interaction — high velocity combined with high rise tends to generate whiffs. |
| Model | Notes |
|---|---|
| Random Forest (baseline) | ranger in R / RandomForestRegressor in Python. Solid baseline, requires imputation for missing values. |
| HistGradientBoostingRegressor (final) | Gradient boosting with histogram binning. Handles Statcast's missing values natively. Faster on 700K+ pitch events. Outperformed RF on both RMSE and R². |
1. Native NaN handling. Statcast data has meaningful missingness — spin_axis isn't recorded for all pitch types, release_extension is occasionally absent. Random Forest requires imputation (which introduces assumptions). HistGradientBoosting learns to route missing values to the optimal branch during training, preserving the signal in missingness itself.
2. Scale. A full MLB season contains 700,000+ pitch events. Histogram binning reduces training time by an order of magnitude compared to Random Forest at this scale.
3. Sequential residual correction. Gradient boosting corrects prior errors iteratively — well-suited to the structured, non-linear relationships between pitch physics and run value outcomes (velocity × break interactions, spin efficiency, etc.).
- Model successfully differentiates elite pitch quality from average — top Stuff+ pitchers align closely with known MLB aces and high-swing-and-miss arms
- Spin-Velocity Ratio ranked among the most predictive engineered features — spin efficiency matters more than raw spin rate
- Induced Vertical Break and release extension were strong predictors across pitch types
- Pitch type was the single most important feature, confirming that evaluating pitches within type context is essential
| Tool | Purpose |
|---|---|
pybaseball |
Statcast data ingestion from Baseball Savant |
scikit-learn |
HistGradientBoostingRegressor, RandomForestRegressor, model evaluation |
pandas / numpy |
Data cleaning, feature engineering |
plotly |
Interactive Stuff+ leaderboard and scatter visualizations |
ranger (R) |
Random Forest baseline implementation in R |
dplyr (R) |
Feature engineering in R pipeline |
git clone https://github.com/jordantaylorkurzweil-hash/stuff-plus-model.git
cd stuff-plus-model
pip install pybaseball scikit-learn pandas numpy matplotlib seaborn plotlyOpen stuff_plus_model.ipynb in Jupyter or Google Colab and run all cells top to bottom.
Note:
pybaseballcaching is enabled. First run pulls ~700K pitch events and may take 2–3 minutes. Subsequent runs use cache.
install.packages(c("ranger", "dplyr"))
# Place training_data.csv and test_data_for_testing.csv in working directory
source("stuff_plus_models.R")- This model is a proxy for Stuff+, not the proprietary Baseball Savant implementation
- Pitcher identity is excluded by design — we want pitch quality, not pitcher reputation
- Count and game context are excluded — Stuff+ is context-neutral by definition
- Trained on 2024 data only — cross-year validation would strengthen generalizability
- Gradient boosting trades interpretability for predictive power; feature importance should be interpreted carefully
This project was built as part of the SABR Analytics Certification (Level IV), which covers advanced modeling techniques applied to baseball data. The Stuff+ metric was developed by analysts at Baseball Savant and is now widely used across MLB organizations to evaluate pitcher arsenals independent of defense and sequencing.
Related work: See also the Baseball Analytics Dashboard repo — a full sabermetrics pipeline covering pitcher WAR, hitter profiles, and team-level analytics using FanGraphs data.
Jordan Kurzweil M.S. Business Administration candidate, Pace University (Lubin School of Business) — August 2026 Google Data Analytics Certificate · SABR Analytics Certification Levels I–IV LinkedIn · GitHub
Data sourced from MLB Statcast via pybaseball. All stats reflect the 2024 MLB regular season.