Abstract
Now more than ever, forecasting the outcomes of U.S. elections is an important and challenging task. Traditionally, statistical or political-science methods have been employed to better understand how individuals will vote. Our approach differs in that we use mathematical modeling. Adapting methods commonly used in epidemiology to understand biological disease transmission, we model the spread of political affiliation (Democratic or Republican) across states using differential equations. We simulate thousands of possible election scenarios, accounting for uncertainty, to make a range of forecasts at the state level. The model’s final forecasts for presidential, senatorial, and gubernatorial elections from 2004 through 2016 have had accuracy comparable to popular forecasting sites, such as FiveThirtyEight. A new focus of our research is on how the accuracy of gubernatorial and senatorial forecasts changes over the months leading up to the election day. We will also discuss our forecasts of the 2020 U.S. elections, which we posted in real time last fall on a website that we created (https://modelingelectiondynamics.gitlab.io/2020-forecasts/). Finally, we will share our work on improving the accuracy of the model by weighting the polling data in different ways. Our research highlights how mathematical modeling can be used for data-driven forecasting on a topic of broad interest and suggests additional research in this field.