Affirm or reverse? using machine learning to help judges write opinions

The US federal court system relies on a system of appellate review where higher courts can either affirm or reverse the decision of the lower courts. We build a machine learning model to predict these appellate affirm/reverse decisions using the text features of the lower-court case under review. The data include all Supreme Court and circuit court decisions since 1880, and most district court decisions since 1923. We use a wide of classification techniques based on n-grams and convolutional neural networks. We achieve an accuracy of 79% in predicting the affirm/reverse decision of circuit courts using district court case text. We achieve 68% accuracy in predicting Supreme Court decisions using circuit court case text.