A groundbreaking machine learning project that predicts Supreme Court voting behavior using historical data from the Supreme Court Database (SCDB). This was the first robust, generalized, and fully predictive model of Supreme Court voting behavior.

Key Achievements

69.7% accuracy in predicting the Court’s overall decisions
70.9% accuracy in predicting individual justice votes
Analyzed 7,700 cases and 68,000+ justice votes from 1953-2013
Covered 30 Justices appointed by 13 presidents across 6 decades

Technical Approach

The model uses an Extremely Randomized Tree method (a variant of random forests) with carefully engineered features based on:

Case characteristics and legal issues
Lower court information
Justice-specific voting patterns
Temporal factors

All predictions use only data available prior to the actual decisions, ensuring the model’s validity for real-world applications.

Impact

This work demonstrated that machine learning could successfully predict judicial behavior at scale, opening new avenues for legal analytics and quantitative legal studies. The project has been widely cited in academic literature and media coverage of computational law.

Publication

A General Approach for Predicting the Behavior of the Supreme Court of the United States
Authors: Daniel Martin Katz, Michael James Bommarito II, Josh Blackman
Published in PLoS One, 2017
Also available on arXiv and SSRN

Note

This repository has been superseded by SCOTUS Predict v2, which extends the analysis to cover 200 years of Supreme Court history.