Forecasting Turnout

We evaluate the predictive power of the leading explanatory models of turnout in the academic literature. We compare the power of using registration, lagged vote, demographics, electoral competition, and early vote data to predict turnout. We specify models to capture each of these approaches to understanding turnout, fit those models to the relevant data from prior elections, and use the estimated parameters from prior years and the relevant observable data from the day of the election in the current year to predict that year’s election. The simplest and most naive model, the Registration Model, out-performed other models in predicting 2016 turnout using 2012 election data and 2020 turnout using 2016 election data. These findings are consistent with classic understandings of which factors most drive turnout, and demonstrate that in modern elections the propensity of registered voters to turnout in presidential elections is fairly stable. Saturated models that combine many of these predictors are common in the academic literature that attempts to explain levels of turnout. We find that such saturated models overfit the data and lead to less accurate predictions than parsimonious models.