Using Modern Neural Networks to Predict the Decisions of Supreme Court of the United States with State-of-the-Art Accuracy

摘要

Deep neural networks are a family of statistical learning models inspired by biological neural networks and are used to estimate functions that can depend on a large number of inputs and are generally unknown. In this paper we build upon the works of Katz, Bommarito and Blackman 2014, who use extremely randomized trees and feature engineering to help in predicting the behaviour of Supreme Court of United States. We explore Machine Learning techniques to achieve our goals including SVM and Neural Networks, but attain state-of-the-art accuracy with Deep Neural Networks trained using momentum methods and incorporating the Dropout technique. We explicitly use only data available prior to the decision and predict the decisions with 70.4 percent accuracy across 7,700 cases with nearly 70,000 justice votes. Our model is simple yet robust, uses far less feature vectors to train and still provides excellent accuracy, but most importantly deploys no feature engineering.