Research Update
November 2, 2020
Election Analytics: Predicting the 2020 Election
Exploratory prediction model for the 2020 Presidential Election provided close predition to the actual Election Day result.
Project Dates: September 2020 - December 2020
Built with: R
Over the final months of the 2020 Presidential Election, I ran an extensive exploration of the 2020 presidential election, along with more traditional election result predictions (fundamentals). This work culminated in a final election prediction model for the 2020 Presidential Election.
Final prediction model overview
The final election prediction model comprised of three parts:
an initial prediction of candidates’ popular vote share based on a weighted model of poll ranking and a decay-over-time effect for each poll’s impact on the final result,
a probabilistic model that predicts the popular vote share using turnout estimates based on the age of the population, the popularity of mail-in and early voting, and the COVID-19 spread in the state, and
an ensemble of the two predictions to correct the probabilistic model’s uncertainty and tie everything back together to produce the final result.
I used two different weight sets for the final ensemble model that generated a conservative and an explorative model, based on how much weight we placed on the probabilistic model compared to the poll ranking and a decay-over-time effect model.
Graphical overview of final prediction model.
Model accuracy
The model proved to be quite reliable in predicting the winner in most states. In particular the conservative model incorrectly predicted the following states: Arizona, Florida, North Carolina, Ohio, and Texas. The explorative model correctly predicted Ohio bringing the number of incorrectly predicted states down to four. You can read an in-depth analysis of the model performance and improvement directions in the post-mortem analysis blog post, and the full analysis in the election analytics blog linked in the Resources section.
Graphical representation of states incorrectly predicted by the explorative final prediction model.
