Predicting U.S. Supreme Court Case Outcomes

Traditional machine learning wisdom requires that we use as much business data as possible when training models. The patterns and insights obtained from that data can then be used to drive business decisions. Unfortunately, not every industry has the dataset or expertise to apply the latest and greatest techniques. This paper attempts to quantify the effect that a sound and specialized methodology can have on the performance of a modest model.

This paper will examine several experiments performed to try and reduce the resources required, dataset size, data preparation, model complexity, and time to train. Once we’ve got a bare bones model, what affect might changing the data preparation or training strategy have on the overall performance? This paper explores 2 novel experiments/approaches to training/testing a model to predict the outcome of U.S. Supreme Court Case Decisions. This paper shows that by grouping cases by issue area, performance can be improved by an average of nearly 4%.