Dem
51
GOP
49
As you may have heard, the Walt Disney Company is downsizing. This is not to please Ron DeSantis, but for normal business reasons. One of the casualties of the downsizing is Nate Silver, who works for the Mouse (albeit indirectly, through its ABC subsidiary). He will be leaving FiveThirtyEight shortly. Apparently, his contract stated that the election model he developed is his personal property and will leave with him, although the details are sketchy. It also appears that his successor at FiveThirtyEight will be another Nate, Nathaniel Rakich, although "Chief Nate" is not a formal position. In any event, Rakich presented the new prediction model yesterday. The model might have been developed due to flaws in the old one, or simply because Silver is taking the old one with him on the way out the door so the site needed a new one.
Rakich didn't go into the details of the model, although in the interest of transparency, there is a webpage discussing it. Briefly, there are separate models for approval ratings, horse-race polls, and other things. The models take into account things like politician favorability, previous elections, generic polls, pollster quality, survey size, flooding the zone, house effects, trendlines, poll recency, polynomial regression, and other factors. The new model uses an exponentially weighted moving average. It also responds to changes more quickly than the old one when multiple pollsters note some change. The model is very complicated and is described in some detail in the link above, if you are interested.
One difference with the old presentation is that error bars are now displayed. There was always a standard deviation, only it wasn't shown. The colored band indicates a probability of 0.95 (2 sigma) that the true result falls in it. Here are the new polls for the Republican primaries:
In 2020, FiveThirtyEight called every state right. We did reasonably well, but not quite as well. We called Arizona, Nevada, Georgia, Wisconsin, Michigan, Pennsylvania, and New Hampshire correctly. Based on an average of the most recent 18 polls, we thought Biden would win Florida and based on an average of 15 polls in the final week, we thought Biden would win North Carolina. The latter was the closest state in the country with Trump ultimately winning by only 1%. That one could easily have gone either way. We also thought Ohio was too close to call, but Trump clearly won it in the end. Both FiveThirtyEight and electoral-vote.com correctly predicted Biden would win. How well the new FiveThirtyEight model will do in 2024 won't be known until after the election. But the trouble with any model is that factors that no one is considering until then, like pollsters not weighting for educational level or reluctance of Trumpy voters to talk to pollsters, can wreak havoc with any model. (V)