An Updated Dynamic Bayesian Forecasting Model for the US Presidential Election

Abstract

During modern general election cycles, information to forecast the electoral outcome is plentiful. So-called fundamentals like economic growth provide information early in the cycle. Trial-heat polls become informative closer to Election Day. Our model builds on and is implemented in Stan . We improve on the estimation of state-level trends, the internal consistency of different predictions at the state and national level, and provide an adjustment for differential nonresponse bias across the cycle. The model forecast a Democratic win with probability in the 80–90% range during most of the 2020 U.S. presidential election campaign, conditional on the two major candidates staying in the race, no major third-party challenges, and no unprecedented challenges with turnout or vote counting.