538 model & poll tracker thread

[Gustaf]

I'm wondering if a reason they waited is that they wanted to know the VP in order to incorporate their (tiny) VP homestate bonus.

[Gustaf]

I'd be inclined to agree if it weren't for the fact that this thread sounded exactly the same 4 years ago and then Silver was right and Atlas was wrong.

[Gustaf]

I'd be inclined to agree if it weren't for the fact that this thread sounded exactly the same 4 years ago and then Silver was right and Atlas was wrong.

The thing is, Silver wasn’t right, he was just the least wrong. And he’s spent four years trying to hold that over everyone else. Given current data, some posters above seem to not understand the distinction between forecasting an event today versus forecasting/predicting the future state of the world. However, I can’t help but feel Silver is trying to cover his a** here a bit in order to claim that his model was best even if it actually misses the mark substantially by using the blackbox of “uncertainty.” In November, say he gives Biden a 3 in 4 chance to win and Biden wins a ‘08 style landslide, he can say “oh we got the winner right but the uncertainty worked out in Biden’s favor.” Or if Trump wins at just a 1 in 4 chance, he can say “the uncertainty went in Trump’s favor, and our model factored that in better than everyone else’s” even though his prediction was technically quite off the mark.

Well, that's why these prediction sites should be evaluated over longer periods of time. 538 did self-evaluation recently and found they were doing pretty well, IIRC.

Of course, this model is different from the previous ones and probably won't ever resurface since it's supposed to be corona-specific so it won't be possible in this case.

[Gustaf]

In the vein of ridiculous 538 maps I'm seeing this one:

[Gustaf]

The model has now been out for a week, and the probabilities haven't really budged.   Biden's polling average lead has gone from 8.2 last Wednesday to 8.5 today, and his probability of winning has only gone from 71% to 72%.  But shouldn't the model be steadily becoming more confident over time even if the polls don't change much?

It’s a reverse-engineered model designed to give Nate Silver cover if he ends up being wrong. I wouldn’t expect much.

Yep.

They said when launching it that polling becomes much more predictable closer to the election. So the model doesn't become steadily more confident, it remains uncertain for a fair amount of time and then rapidly gets more certain as we're closer to the election.

[Gustaf]

Can someone explain how the odds dropped from Biden 73/100 to 70/100 overnight?

Biden's nat'l average literally hit +9.3 yesterday and is down to +8.8, but only b/c of Ipsos (7) and Change (, but those are still on par with recent.

Only state polls I've seen are Trafalgar, and the Change polls which were 4/6 better than two weeks ago.

I'm finding it odd how when Biden gets good polls, the odds barely budge up, but then if there's like 1 or 2 bad (not even really that bad) polls, it drops a whole 3%? What?

Doesn't the model have some sort of convention bounce effect? That might affect it. It might think Biden should be on a convention bounce high right now or something.

[Gustaf]

IIRC, it includes national polling by creating a trend line. So if Biden loses a point from July nationally they adjust the state polls by a point (not exactly, because there is elasticity etc but basically)

[Gustaf]

And we've gotten Biden +12, Biden +13, and Biden +9 this morning, and it's still at Biden +8.0.

Once again, for whatever reason, it's as if good Biden polls don't help all that much, but a bad one hurts way more than the good ones help.

There was two new PA polls today with Biden +9. There was also a Biden +7 poll, Nate has Biden only up  4.7 in PA.

So, in spite of some favorable polls, Biden's lead went from 5.4 on Sunday to 4.6 today, mainly due to the national lead shrinking from 8 to 6.9. Shouldn't state polls cancel that out?

Also, Nate tried to say before that national polls don't have that much weight, but they clearly do. Even so, shouldn't that effect the *forecast* though? His state averages model should literally just be *averages* without any other f**kery going on. I want an RCP aggregator but just not sh**tty.

As far as I've understood there are no state averages.

There is a national polling average which is kind of its own thing. Then there are state projections based on polls and fundamentals for that state. There is also a national forecast that is created by aggregating the state forecasts.

The only role the national average plays is that it affects the impact of state polls. So Biden's current lead in Pennsylvania is weighted by polls there from say the last month and some fundamentals factor. If his national lead has gone down compared to 1 month ago his lead in a poll from a month ago is also adjusted down.

How big of an impact that is depends a bit, obviously. In general I'd guess a month old poll won't have much impact anyway, but now that we don't have that many polls in general I'm guessing it's more sensitive to those things.

[Gustaf]

538 has Biden up to 74% now, highest since July.

[Gustaf]

If you hover over the states in the snake chart you can now see the projected final percentages. Praise!

[Gustaf]

It’s getting a little bit creepy how stable the model is right now.  Biden has been at 77 for 7 straight days.  It seems like there should be a little movement just by random chance.  I’m surprised Nate isn’t tweeting worries about herding yet.

The model assumes pretty strong convention bounces that last for up to a month.  Next week could be significant as I believe that's when Trump's assumed bounce finally goes to 0.  The narrowing to Biden being only a 2:1 favorite in late August was because the model assumed a DNC polling bounce that didn't really materialize.  I have a feeling it's overweighting the RNC as well and perhaps giving Biden too much credit right now.

I don't think the bounces are meant to last that long, where have you seen that? The tightening in mid-August was because Trump was getting an RNC bounce and Biden wasn't getting a DNC bounce.

538 has written several articles on how they incorporate convention bounces into previous iterations of the model and how they expect them to last 2-3 weeks if not interrupted by another party's convention. If you look at the "nowcast" element of their model (they're not calling it that, but it exists, the adjustments made to Trump's and Biden's polling numbers seem to come from "current events" which I would assume is them attempting to measure the remainders of the convention bounces in the polls they're using.

Adjustments for the RNC bounce will disappear on a state-by-state basis, depending on how quickly state-level polls come out.

They seem to currently be given Biden 0.1-0.2% extra and deflating Trump by like 0.4-0.5% so if polling stays the same Trump should improve a little bit in the model once that rolls off. On the other hand, Trump is getting a lot from the fundamentals-projection part of the model and that should be diminishing fairly quickly by now as we approach election day. So overall if polling doesn't move Trump's chances should be going down.

[Gustaf]

Nate made a table somewhere with the exact percentage for the fundamentals. It's falling towards 0, which it will be on election day.

It doesn't fall linearly, it falls faster and faster as you approach election day.

[Gustaf]

Ah, I found it: https://fivethirtyeight.com/features/how-fivethirtyeights-2020-presidential-forecast-works-and-whats-different-because-of-covid-19/

It was 16% 50 days out, will be 11% 25 days out, 6% 10 days out and 3% 5 days out.

[Gustaf]

0.4% popular vote gap in Georgia now, closest it's ever been

Again, that damn 'Incumbency and Economics' caveat

The "Incumbency and Economics" factor is falling fast now. I think Silver said recently that Trump would be at 5% if the election were now with these polls.

[Gustaf]

Trump is now down to 87-13.

[Gustaf]

Georgia has flipped back to Biden and he is now at his all- time high likelihood of winning there at 52%.

Iowa has now flipped back to Biden as well.  Though I'm surprised all the great state polls Biden has gotten today haven't really impacted his national win probability.

Looking at recent days fluctuation I think it was on the verge of 87/88 and might have moved towards more solid 88 territory. Flipping close states from 49-51 to 51-49 also won't impact a probabilistic model that much.

[Gustaf]

Lol, 538 isn't considering Wisconsin and Michigan to be competitive races anymore.

[Gustaf]

Why are people engaging with a moron who has shown no understanding of statistics (or anything) and has been hilariously wrong election after election with his own mind-models?

[Gustaf]

Why are people engaging with a moron who has shown no understanding of statistics (or anything) and has been hilariously wrong election after election with his own mind-models?

Being in argument with someone in the minor leagues can be therapeutic as well as good practice for explaining things IRL.

I'd agree it can be fun, but people should be clear about how no one takes this clown seriously.