Author ORCID Identifier

https://orcid.org/0009-0001-0887-0890

Document Type

Thesis

Date of Award

2026

Degree Name

Master of Science (MS)

Department

Business

First Advisor

Bartlomiej Hanus

Abstract

Federal appellate courts are the final arbiters in many cases, yet systematic machine learning analysis across all twelve circuits remains largely absent from the computational law literature. With courts of appeals deciding tens of thousands of cases annually and the Supreme Court reviewing only a fraction, understanding what predicts reversal outcomes has both theoretical importance and practical consequences for litigants, attorneys, and judicial administrators. This study addresses that gap using eleven years of federal appellate decisions from the Federal Judicial Center’s Integrated Database. A systematic comparison of twenty-five machine learning models, spanning five algorithms and five class-imbalance correction strategies, identifies the strongest predictors of appellate reversal. The best-performing model is validated through temporal holdout testing and bootstrap resampling to confirm generalizability beyond the training period. The study also examines preliminary patterns following Loper Bright Enterprises v. Raimondo (2024), which eliminated forty years of Chevron deference. This research extends the computational law tradition to the intermediate appellate level, contributing to a replicable multi-circuit analytical framework and a systematic treatment of class imbalance in legal prediction.

Subject Categories

Law | Statistics and Probability

Keywords

Access to justice Federal appellate courts Judicial outcome prediction Machine learning Pro se litigants XGBoost

Number of Pages

89

Publisher

University of South Dakota

Available for download on Sunday, July 26, 2026

Share

COinS