Under-performing, over-performing, or just performing? The limitations of fundamentals-based presidential election forecasting

U.S. presidential election forecasts are of widespread interest to political commentators, campaign strategists, research scientists, and the public. We argue that most fundamentals-based political science forecasts overstate what historical political and economic factors can tell us about the probable outcome of a forthcoming presidential election. Existing approaches generally overlook the uncertainty in coefficient estimates, decisions about model specifications, and the translation from popular vote shares to Electoral College outcomes. We introduce a Bayesian forecasting model for state-level presidential elections that accounts for each of these sources of error, and allows for the inclusion of structural predictors at both the national and state levels. Applying the model to presidential election data from 1952 to 2012, we demonstrate that, for covariates with typical levels of predictive power, the 95% prediction intervals for presidential vote shares should span approximately ±10% at the state level and ±7% at the national level.