|
Abstract:
|
This work extends commonly used Bayesian methods in election forecasting to simultaneously estimate polling errors and incorporate those estimates into election forecasting. We develop a flexible model for polls and the election outcome as a peak at an underlying hidden Markov process. Polls are modeled as noisy measurements of the electorate’s current preferences, which evolve over time, and the true preferences are revealed on election day. The final reveal allows estimation of systematic error in the measurements of the underlying process while accounting for temporal variability. Then, the same model is used to predict the outcome of future elections while accounting for measurement error learned from past results. This fusion of error estimation and forecasting fully captures uncertainty surrounding future elections, which is crucial for campaigns allocating limited resources to state-level contests. Results show that the 2016 US Presidential election was an unusually large polling error, and accounting for that error in 2020 polling data results in much more conservative forecasts of Biden’s chances of victory than other analyst’s predictions.
|
