Using Machine Learning Techniques to Predict United States House of Representatives Elections

Abstract

We predict the results of the United States House of Representatives elections using machine learning techniques. We started by collecting and preprocessing data on the partisan lean of districts, the state of the economy, the national political environment, candidate political stances, news headlines about each candidate in each race, and the past election results. Then, we selected, designed, and trained the models we would use to predict those election results. We used single-tier models that took either only news headline text data or only numerical data as inputs and two-tier models that used both news headlines and numerical data. Our best-performing model was a two-tier model with a GRU as the first tier followed by a Ridge Regressor as the second tier, with a root mean squared error of under 2 percentage points. The vote share predicted by our best model was within 2 percentage points of the actual observed vote share.