Election polling is not dead: A Bayesian bootstrap method yields accurate forecasts

We present a new Bayesian bootstrap method for election forecasts that combines traditional polling questions about people’s own intentions with their expectations about how others will vote. It treats each participant’s election winner expectation as an optimal Bayesian forecast given private and public evidence available to that individual. It then infers the independent evidence and aggregates it across participants. The bootstrap forecast outperforms aggregate national polls in the 2020 U.S. election, as well as the forecasts based on traditional polling questions posed on large national probabilistic samples before the 2018 and 2020 U.S. elections. The bootstrap forecast puts most weight on people’s expectations about how their social contacts will vote, which might incorporate information about voters who are difficult to reach or who hide their true intentions. Beyond election polling, the new method is expected to improve the validity of other social science surveys.

SIMPLE BAYESIAN FORECAST COMBINATION

In this paper, it is proposed to combine the forecasts using a simple Bayesian forecast combination algorithm. The algorithm is applied to forecasts from three non-nested diffusion models for S shaped processes like virus diffusion. An illustration to daily data on first-wave cumulative Covid-19 cases in the Netherlands shows the ease of use of the algorithm and the accuracy of the newly combined forecasts.

Forecasting virus diffusion with simple Bayesian forecast combination

There are various diffusion models for S shaped processes like virus diffusion and these models are typically not nested. In this note it is proposed to combine the forecasts using a simple Bayesian forecast combination algorithm. An illustration to daily data on cumulative Covid-19 cases in the Netherlands shows the ease of use of the algorithm and the accuracy of the thus combined forecasts.