Data & AnalysisBeta

Compare Lessons Glossary About

Leave a Note

Bug reports, feature ideas, questions -- all welcome. No account needed.

10 more characters needed

This app is entirely vibe-coded. Leave a note if you find something broken.

Data powered by Statcast

·About·Support·

Built by Robert Stock

BetaNot affiliated with or endorsed by Major League Baseball. Stuff+, Location+, and Pitching+ are Stockyard models -- not the Eno Sarris variants or official MLB statistics.

How Models Work

Overview Model Stack Boosted Trees & Early Stopping Tree Depth & Command Location+Stuff+Pitching+Arsenal Synergy Projections Data & Features Limitations & FAQ

How Models Work

A guided tour of the modeling stack.

A quick map of the model stack and how the pieces fit together.

The pipeline order: Location+ -> Stuff+ residuals -> Pitching+.

Boosted Trees & Early Stopping

Why boosting works and how early stopping protects generalization.

Tree Depth & Command

A case study in underfitting and why one extra split matters.

Command-only model that grades where the pitch crosses the plate.

Physics-only model trained on residual xRV after Location+.

Combined model that merges command and stuff into one grade.

Arsenal Synergy

How individual pitch grades add up to a cohesive arsenal.

Next-season outcomes using historical skill signals.

Data & Features

What data powers the models and how features are built.

Limitations & FAQ

Where the models are strong, and where they should be ignored.

Boosted Trees & Early Stopping

Model Stack

The full pipeline from pitch data to grades

The stack is intentionally layered. Command is modeled first, then stuff learns on residuals, and Pitching+ reunites the full signal. This page walks through each stage and explains why the order matters.