Virginia Mega Thread: The Youngkin Administration

[JustinSmith]

My model has the new FOX News poll taking off about 12,000 votes from McAuliffe and adding about 14,000 votes to Youngkin. But it's still overwhelmingly in favor of a McAuliffe win.

[JustinSmith]

I think a lot of election prediction models (including mine) may stumble over this one. I've always felt that a prediction model's worth is based on how far ahead of the election it is accurate. I don't think making a correct prediction the day before the election is that valuable. This election has followed the difficult-to-predict sequence of being extremely consistent in favoring one candidate for months, with a sudden swing in the last few days. Right now, nobody has a model that accounts for this happening months before it happens. My own model is still predicting McAuliffe at +5%, but I think it may have a weakness when it comes to stark polling changes at the end of the election cycle. Part of the difficulty in trying to build a model that can be accurate months ahead of time is that it has to be resistant to the small day-to-day changes. An eventual goal of such a model should be to recognize when those small day-to-day changes are meaningful.

[JustinSmith]

I think a lot of election prediction models (including mine) may stumble over this one. I've always felt that a prediction model's worth is based on how far ahead of the election it is accurate. I don't think making a correct prediction the day before the election is that valuable. This election has followed the difficult-to-predict sequence of being extremely consistent in favoring one candidate for months, with a sudden swing in the last few days. Right now, nobody has a model that accounts for this happening months before it happens. My own model is still predicting McAuliffe at +5%, but I think it may have a weakness when it comes to stark polling changes at the end of the election cycle. Part of the difficulty in trying to build a model that can be accurate months ahead of time is that it has to be resistant to the small day-to-day changes. An eventual goal of such a model should be to recognize when those small day-to-day changes are meaningful.

Is there tangible proof though that Youngkin has the "momentum"?

No, which is why my model is still +5% for McAuliffe. My other comments are based on what "feels right" that fly in the face of what my model says.

[JustinSmith]

Also if the polls are off on this race too, I think American polling is fundamentally useless going forward after the debacles over the last few cycles.

Not necessarily, if you can identify a pattern as to the way in which they're off, they can still be useful. A clock that is 10 minutes ahead can be used to determine the time if you know it's 10 minutes ahead.

[JustinSmith]

If you go on the FEC Website and type in Glenn Miller in McLean, who's a lawyer, he's donated thousands of dollars to Trump Republicans, including Kelly Loeffler, Kim Klacik, and Lindsay Graham (2020), Ted Cruz (2018), among others.  Not sure he really cares about someone being a "Trump Republican."  Youngkin may well win, but not sure Mr. Miller was ever really a getable Democratic vote.

I keep seeing this kind of thing happening. Someone claims "I used to X Party but the last guy was really terrible and now I'm voting for Y Party" and then it comes out they were never X Party to begin with.

What is the point of lying in this kind of way? I don't understand that there is any tangible benefit to helping the candidate you want to see win by pretending you used to support the other side. Sure, it might make people in the other party "get worried" but if anything that's more likely to drive them out to vote more.

[JustinSmith]

I think that any claim that a candidate won or lost because of one single thing is, in most cases, going to be very wrong. Populations of millions of people are pretty complex structures and it's highly unlikely that everyone has the same motivations for why they vote the way they do.

It's certainly looking like my prediction evaluation for this election is going to be well into what I've defined as the "unacceptable" category. Which I means I should have plenty of meat to study for improving the model. It's easy to see what happened on a basic level. Although the model favored recent polls more than older polls, the sheer number of polls that had been done prior to the sudden shift still managed to outweigh the newer ones.

I've felt for some time that an election prediction model is valuable only if it can offer an accurate prediction far in advance of the election. An accurate same-day prediction isn't really that useful, except perhaps for the last minute gamblers. The real puzzle to solve is how to anticipate a shift like the one we saw for VA. A nobel prize for the prediction model-maker who can devise the algorithm capable of predicting the October VA shift back in June.

In that regard, there's an interesting question to consider: Did the shift towards the end of the polling cycle reflect voters changing their minds about who they would vote for, or did it reflect the early polls being inaccurate regarding how voters at that time planned to vote? It feels like an actual change, but I'm not really sure that any actual data exists to answer this question definitively.

[JustinSmith]

Ciattarelli has just taken the lead in NJ.