Our work is based on a model by Alexandria Volkening, Daniel F. Linder, Mason A. Porter, and Grzegorz A. Rempala (to appear in SIAM Review). The model takes a SIS compartmental approach and accounts for interactions between voters in different states. Adapting ideas from mathematical biology, we study the evolution of the percentages of Democratic, Republican, and undecided/other voters in time. We use polling data from FiveThirtyEight
that we average by month to determine the parameters in our model, and we use the model to simulate 10,000 elections with demographics-correlated noise to generate our forecasts.
All of the code to reproduce or build on our forecasts is available here.
To forecast the 2020 races, we build on an SIS compartmental modeling approach that was previously applied to U.S. gubernatorial, senatorial, and presidential elections in 2012, 2016, and 2018. To determine our model parameters, we use polling data from FiveThirtyEight. To determine the states that appear in our safe red and safe blue superstates, we utilize the ratings by 270towin (the consensus version) and FiveThirtyEight. We include correlated noise in our model based on state-level demographic information from the U.S. Census Bureau. We provide our references below: