Improving Supreme Court Forecasting Using Boosted Decision Trees

Citation:

Kaufman, Aaron, Peter Kraft, and Maya Sen. Submitted. “Improving Supreme Court Forecasting Using Boosted Decision Trees”. Copy at http://j.mp/2nRjto6

Abstract:

What predicts the behavior of Justices on the U.S. Supreme Court? Previous attempts to develop predictive models of Supreme Court behavior leave room for improvement, in large part because of challenges involved with incorporating all available case-specific factors. In this article, we address these issues using an AdaBoost decision tree regressor, a popular approach in machine learning that is relatively underused in political science. We couple this approach with a novel mixed data set of both oral arguments data as well as data on case-level attributes. As we show, our AdaBoosted approach substantially outperforms not only existing predictive models of Supreme Court outcomes but also the predictions of legal experts. Substantively, this suggests that combining both legal information and the information revealed by the Justices themselves in the months leading to the decision provide the best information. We conclude the article by discussing possible applications of the AdaBoost approach within the social sciences.

Last updated on 02/14/2017