US 2020 Presidential Election prediction thread

[DGuller]

Isn't 9% their "2020 fudge factor"? So they're basically saying that there is nothing in the data at all that gives Trump a chance in hell, it's only down to unknown unknowns.

[alfred russel]

« I am here to remind you Trump can still win »

https://fivethirtyeight.com/features/im-here-to-remind-you-that-trump-can-still-win/amp/?__twitter_impression=true

"My career is based on projecting winners - and primarily presidential races which is why my website is called 538. I think Biden will win but I'm not sure so please let my career continue if Trump wins."

[FunkMonk]

Nate S. has no balls. He's been hedging for last the month or so on Twitter.

"Yes, Biden is the favorite. BUT here is why Trump may win."

I think everyone clearly understands this is a definite possibility, Nate. It's why some polling showed that people expected Trump to win the election even though he's been seriously behind the entire time. It is the year 2020, not 2016. Everyone knows Trump can still win.

[Sheilbh]

"My career is based on projecting winners - and primarily presidential races which is why my website is called 538. I think Biden will win but I'm not sure so please let my career continue if Trump wins."

I'm aware I'm not a person with enough statistical knowledge to make a real critique. But it's something I've never really got with 538. Like there's a 90% chance Biden wins and if Biden wins, the model works. And there's a 10% chance Trump wins and if Trump wins, the model works. I don't really get the point. Like the whole model thing doesn't seem to add anything to me.

I find the stuff about how they weigh polls etc and the technical stuff on polls quite interesting. And I think they're pretty good about the electoral college paths. But I've never really understood the whole 538/Nate Silver thing

[Caliga]

Nate S. has no balls. He's been hedging for last the month or so on Twitter.

Not surprising, since a) this is the guy's career, and b) he was wrong in 2016, so he's got to be nervous as fuck about being wrong again.

[FunkMonk]

The guy who runs the Economist's model has been pretty bullish on Biden for a while now. Their model shows the probability of Trump winning at 5%.

[DGuller]

Nate S. has no balls. He's been hedging for last the month or so on Twitter.

Not surprising, since a) this is the guy's career, and b) he was wrong in 2016, so he's got to be nervous as fuck about being wrong again.

It's very likely that he wasn't wrong in 2016, though it is unknowable.  Inside straight draws do hit sometimes, that doesn't make the properly calculated probabilities of them hitting wrong.  You're wrong if your 70% probability events always happen.  In fact, it's possible that he even gave Trump too much of a probability, maybe Trump needed to get more lucky than hitting a 30% shot.

[DGuller]

I'm aware I'm not a person with enough statistical knowledge to make a real critique. But it's something I've never really got with 538. Like there's a 90% chance Biden wins and if Biden wins, the model works. And there's a 10% chance Trump wins and if Trump wins, the model works. I don't really get the point. Like the whole model thing doesn't seem to add anything to me.

I find the stuff about how they weigh polls etc and the technical stuff on polls quite interesting. And I think they're pretty good about the electoral college paths. But I've never really understood the whole 538/Nate Silver thing

The problem is that you can't evaluate a single prediction, it's simply impossible.  You can only evaluate someone's record in aggregate.  If Nate Silver makes 1000 90%/10% predictions, and the 90% predicted event happens 900 times out of 1000, then overall he has his probabilities well-calibrated.  You can't tell anything at all about a single 90%/10% prediction, regardless of how it comes out.  That doesn't mean that you shouldn't try to assign probabilities to events, in fact it would be a big mistake not to, you just have to understand that you can't validate these probabilities by how reality plays out.

[Maladict]

As I mentioned earlier, I think the correlation assumptions might be a bit too strong here.  If PA goes Trump because Republicans are successful with their funny business, that wouldn't indicate much for the other states.  The correlations would make sense if PA goes Trump because the polls missed some dynamic systemically.

Funny business is not part of the model, so yes it would assume a structural shift among the voters that the polls missed.

[FunkMonk]

"My career is based on projecting winners - and primarily presidential races which is why my website is called 538. I think Biden will win but I'm not sure so please let my career continue if Trump wins."

I'm aware I'm not a person with enough statistical knowledge to make a real critique. But it's something I've never really got with 538. Like there's a 90% chance Biden wins and if Biden wins, the model works. And there's a 10% chance Trump wins and if Trump wins, the model works. I don't really get the point. Like the whole model thing doesn't seem to add anything to me.

I find the stuff about how they weigh polls etc and the technical stuff on polls quite interesting. And I think they're pretty good about the electoral college paths. But I've never really understood the whole 538/Nate Silver thing

I think, for the general audience, people really just like seeing the numbers go up and down for their candidate or against their opponent. Nate has made a career out of providing people with the opportunity to psyche themselves up or depress themselves into oblivion.

For me, these models help frame the structure of probabilities for the outcome of the election, but I don't take what they say as gospel. I think 538's poll aggregator is much more practical.

[DGuller]

I personally don't take 538 probabilities as a gospel either, but I do take them as the best available answer.  Model error and model uncertainty definitely are things, but sometimes you just don't have a better alternative than to go with them, you don't always have the luxury of waiting to form your opinion until a gospel model arrives at your door.

[The Brain]

I personally don't take 538 probabilities as a gospel either, but I do take them as the best available answer.  Model error and model uncertainty definitely are things, but sometimes you just don't have a better alternative than to go with them, you don't always have the luxury of waiting to form your opinion until a gospel model arrives at your door.

Give it a week or two.

[Sheilbh]

The problem is that you can't evaluate a single prediction, it's simply impossible.  You can only evaluate someone's record in aggregate.  If Nate Silver makes 1000 90%/10% predictions, and the 90% predicted event happens 900 times out of 1000, then overall he has his probabilities well-calibrated.  You can't tell anything at all about a single 90%/10% prediction, regardless of how it comes out.  That doesn't mean that you shouldn't try to assign probabilities to events, in fact it would be a big mistake not to, you just have to understand that you can't validate these probabilities by how reality plays out.

Right but I suppose I just don't get the value of assigining probabilities if you're not actually working on one of the campaigns - especially because, from memory, Silver started because he was annoyed at horse race reporting.

I think it's useful to have someone who can with real methodology assign values to polls because that helps us understand them, which is important because they're a really big bit of reporting on elections. But otherwise it just feels like Silver's actually ended up just replacing the horse race people are reporting on.

I think, for the general audience, people really just like seeing the numbers go up and down for their candidate or against their opponent. Nate has made a career out of providing people with the opportunity to psyche themselves up or depress themselves into oblivion.

That's probably it - I should say I have really enjoyed the wildly, crazily wrong models in UK elections. I think they've actually stopped doing them after the debacles of 2010 and 2015. So in 2017 and 2019 they've just done pieces after the event about how the results aren't that surprising

For me, these models help frame the structure of probabilities for the outcome of the election, but I don't take what they say as gospel. I think 538's poll aggregator is much more practical.

Yeah - I agree. Those are the things I find useful - they frame paths and the poll reporting is really useful.

[alfred russel]

It's very likely that he wasn't wrong in 2016, though it is unknowable.  Inside straight draws do hit sometimes, that doesn't make the properly calculated probabilities of them hitting wrong.  You're wrong if your 70% probability events always happen.  In fact, it's possible that he even gave Trump too much of a probability, maybe Trump needed to get more lucky than hitting a 30% shot.

This may be a matter of semantics, but I think everyone would agree that on election day 2016 Trump didn't have a 30% chance of winning--he had a ~100% chance. There were ~140 million americans that had voted or intended to vote, and the right number of those in the right places intended to vote for Trump such that Trump would be elected president.

Of course no one knew which 140 million Americans would vote, and what each of them was thinking. Hence we use polls, which provide imperfect information of what the general public is thinking, which gets put into models, which have their own simplifying assumptions and sources of error.

So when Nate Silver gave the precise estimate that Trump had a 28.6% chance, what we would really saying was that his model projected that Hillary would win, but because the model is imperfect and was also fed with imperfect data, he thought the model would produce the wrong answer 28.6% of the time.

I understand the marketing angle that would cause Silver to prefer it expressed as "trump has a 28.6% chance of winning" versus "there is a 28.6% chance my model is wrong", but I think the die is more or less cast at this point and what is going to happen will happen, and if you threw more resources at this question you could arrive at a more precise estimate.

[DGuller]

This may be a matter of semantics, but I think everyone would agree that on election day 2016 Trump didn't have a 30% chance of winning--he had a ~100% chance.

That's wrong.  I know I definitely wouldn't agree, so no, not everyone would agree.  What you're saying is that the probability of pulling an ace out of a shuffled deck is not 1/13, but either 0 or 1.  The deck is already shuffled, so the top card is predetermined, so it's either an ace or not, right?

What lay people often fail to grasp, because it is pretty tricky and unintuitive to grasp, is that uncertainty due to imperfect information is for all purposes exactly the same as the uncertainty due to a truly random effect.  The probability of pulling an ace is the same before shuffling and after shuffling, it's still 1/13.  We had no way of knowing how the people who were going to vote would vote, even if it was predetermined, like a deck is after shuffling.  Unknowably predetermined is no different from unknown.