538 model & poll tracker thread

[Brittain33]

This is just embarrassing.

40,000 simulations or not, this shouldn't show up if you do 40,000,000 simulations.

Titanium Tilt R Florida and "Just Not There Yet" Georgia.

Titanium R Oregon. Fivey Fox took drugs again.

That one poll showing Oregon as a 12-point gap has polluted the data.

[Astatine]

This is just embarrassing.

40,000 simulations or not, this shouldn't show up if you do 40,000,000 simulations.

Titanium Tilt R Florida and "Just Not There Yet" Georgia.

Titanium R Oregon. Fivey Fox took drugs again.

That one poll showing Oregon as a 12-point gap has polluted the data.

538 was obsessed with red Oregon even before... Just take a look at the thread "weird 538 maps".

[Figueira]

I've noticed that whenever there's a rounding area causing the percentages to add up to 99 instead of 100, they include a tied map in their 100 randomly selected maps. Seems a little misleading.

[Fmr. Gov. NickG]

I've noticed that whenever there's a rounding area causing the percentages to add up to 99 instead of 100, they include a tied map in their 100 randomly selected maps. Seems a little misleading.

Isn't a tied map the only reason the %'s wouldn't add up to 100?

I assume that the model isn't literally taking 100 random maps from among all the maps.  If Biden is projected to win 77% of the time, it must always choose exactly 77 maps where Biden is projected to win.  Otherwise, the composition of this sample would fluctuate enormously, as the margin of error for a sample of 100 is around 10%.

Given this, is seems reasonable to deliberately choose exactly one tied map when the tie possibility makes a meaningful difference in the percentages.

[Figueira]

I've noticed that whenever there's a rounding area causing the percentages to add up to 99 instead of 100, they include a tied map in their 100 randomly selected maps. Seems a little misleading.

Isn't a tied map the only reason the %'s wouldn't add up to 100?

I assume that the model isn't literally taking 100 random maps from among all the maps.  If Biden is projected to win 77% of the time, it must always choose exactly 77 maps where Biden is projected to win.  Otherwise, the composition of this sample would fluctuate enormously, as the margin of error for a sample of 100 is around 10%.

Given this, is seems reasonable to deliberately choose exactly one tied map when the tie possibility makes a meaningful difference in the percentages.

Technically yes, but just because the numbers currently don't add up to 100 doesn't mean that the odds of a tie have increased.

And yeah, they are deliberately choosing a number of maps that reflects the current percentage for each candidate. I'm not sure how the algorithm for that works though.

[Crumpets]

I looked at the GA details and the current "now-cast" in Georgia is a Trump win of just 0.4%, while the economics-incumbency model has Trump leading by a full 9.6%. The former makes up 87% of the projected share while the latter makes up 13%. Watch for Georgia's probability numbers start to converge on 50/50 if polling stays where it's currently at.

[Crumpets]

Another fun fact from the current projection: 538 currently has Alaska as more likely to go for Biden (21%) than Wisconsin for Trump (19%).

[Gass3268]

I looked at the GA details and the current "now-cast" in Georgia is a Trump win of just 0.4%, while the economics-incumbency model has Trump leading by a full 9.6%. The former makes up 87% of the projected share while the latter makes up 13%. Watch for Georgia's probability numbers start to converge on 50/50 if polling stays where it's currently at.

How do you get the now cast?

[Crumpets]

I looked at the GA details and the current "now-cast" in Georgia is a Trump win of just 0.4%, while the economics-incumbency model has Trump leading by a full 9.6%. The former makes up 87% of the projected share while the latter makes up 13%. Watch for Georgia's probability numbers start to converge on 50/50 if polling stays where it's currently at.

How do you get the now cast?

If you go to each state page and scroll down to the "How do we get from polls to forecasted vote share?" section, there's an entry for "Today's vote share," which I believe is the same number they used for their "now cast" in the past, even if it looks like they've retired the name.

[GeorgiaModerate]

https://twitter.com/NateSilver538/status/1311369742951870464

[Antonio the Sixth]

I've noticed that whenever there's a rounding area causing the percentages to add up to 99 instead of 100, they include a tied map in their 100 randomly selected maps. Seems a little misleading.

Isn't a tied map the only reason the %'s wouldn't add up to 100?

I assume that the model isn't literally taking 100 random maps from among all the maps.  If Biden is projected to win 77% of the time, it must always choose exactly 77 maps where Biden is projected to win.  Otherwise, the composition of this sample would fluctuate enormously, as the margin of error for a sample of 100 is around 10%.

Given this, is seems reasonable to deliberately choose exactly one tied map when the tie possibility makes a meaningful difference in the percentages.

Technically yes, but just because the numbers currently don't add up to 100 doesn't mean that the odds of a tie have increased.

And yeah, they are deliberately choosing a number of maps that reflects the current percentage for each candidate. I'm not sure how the algorithm for that works though.

Well, they estimate the probability of a tie to be around 0.5%, so if you're using 100 dots it seems fair to include it about half the time.

[Crumpets]

Here's an interesting tidbit from their latest write-up:

As I wrote earlier in the piece, our forecast gives Trump about a 9 percent chance of winning an election held today despite his current deficit in polls — not bad when you’re 7 points down! But it’s about equally likely that a 7-point Biden lead could translate into a 12-point Biden win, in which he’d not only carry states like Georgia and Texas, but would also have a shot in South Carolina, Alaska and Montana.

https://fivethirtyeight.com/features/trumps-chances-are-dwindling-that-could-make-him-dangerous/

[Fmr. Gov. NickG]

This model has become completely perplexing to me.  

On Sept. 19, 11 days ago, Biden was 77% to win.
During the past 11 days, the following things have happened:
1.) Biden’s national lead has grown a full percentage point, from 6.6 to 7.6
2.) There have been three A or A+ quality polls of the tipping point state (PA) suggesting Biden is winning the state by more than his national lead
3.) The election has gone from being 45 days away to 34 days away, which should substantially reduce uncertainty.

And yet Biden’s win % has only increases a single point, to 78%.
How is this possible?  What is actually necessary to move the needle?

[Antonio the Sixth]

This model has become completely perplexing to me. 

On Sept. 19, 11 days ago, Biden was 77% to win.
During the past 11 days, the following things have happened:
1.) Biden’s national lead has grown a full percentage point, from 6.6 to 7.6
2.) There have been three A or A+ quality polls of the tipping point state (PA) suggesting Biden is winning the state by more than his national lead
3.) The election has gone from being 45 days away to 34 days away, which should substantially reduce uncertainty.

And yet Biden’s win % has only increases a single point, to 78%.
How is this possible?  What is actually necessary to move the needle?

The main issue seems to be that Biden was being artificially propped up because the model thought the RNC convention bounce was still going on at that point (kind of weird that they modeled such a long convention bounce, but I guess they're basing themselves off previous cycles when this stuff had more of an impact). Now, that bounce adjustment has almost entirely worn off, which has had the effect of blunting Biden's gains.

Even then, these gains are far from uniform. He's continued to get bad polls in FL, NC and AZ, which while not the most likely tipping points, would be a major freiwall for Biden if he could pull away there.

[emailking]

How is this possible?  What is actually necessary to move the needle?

He's going to gain a point every 3 or 4 days from now until election night. Unless the bottom falls out on Trump. That's what will happen.

[Smash255]

This model has become completely perplexing to me. 

On Sept. 19, 11 days ago, Biden was 77% to win.
During the past 11 days, the following things have happened:
1.) Biden’s national lead has grown a full percentage point, from 6.6 to 7.6
2.) There have been three A or A+ quality polls of the tipping point state (PA) suggesting Biden is winning the state by more than his national lead
3.) The election has gone from being 45 days away to 34 days away, which should substantially reduce uncertainty.

And yet Biden’s win % has only increases a single point, to 78%.
How is this possible?  What is actually necessary to move the needle?

The main issue seems to be that Biden was being artificially propped up because the model thought the RNC convention bounce was still going on at that point (kind of weird that they modeled such a long convention bounce, but I guess they're basing themselves off previous cycles when this stuff had more of an impact). Now, that bounce adjustment has almost entirely worn off, which has had the effect of blunting Biden's gains.

Even then, these gains are far from uniform. He's continued to get bad polls in FL, NC and AZ, which while not the most likely tipping points, would be a major freiwall for Biden if he could pull away there.

In the model they did state that the Convention bounce they were adding were going to be less than normal, though the GOP one would last longer due to the timing of it being second.  Also even if the bounce itself was 10-15 days or so keep in mind the polls in with the average will be a bit longer than that.

[Roll Roons]

https://projects.fivethirtyeight.com/2020-election-forecast/

As of this morning, Biden now has a 79% chance.

[Crumpets]

https://projects.fivethirtyeight.com/2020-election-forecast/

As of this morning, Biden now has a 79% chance.

Looks like it's up to 80 now, at least as I'm looking at it. No jaw-dropping new polls since yesterday - maybe now that the first debate is over and there wasn't a clear Trump win they have uncertainty going down?

[CellarDoor]

https://projects.fivethirtyeight.com/2020-election-forecast/

As of this morning, Biden now has a 79% chance.

Looks like it's up to 80 now, at least as I'm looking at it. No jaw-dropping new polls since yesterday - maybe now that the first debate is over and there wasn't a clear Trump win they have uncertainty going down?

Probably.  We're also at the point where the limited time remaining becomes a larger factor in their model.  There just isn't much time to make up a lot of ground, especially considering that people are already voting. 

[American2020]

https://twitter.com/FiveThirtyEight/status/1311675953026924546

[Figueira]

https://projects.fivethirtyeight.com/2020-election-forecast/

As of this morning, Biden now has a 79% chance.

Looks like it's up to 80 now, at least as I'm looking at it. No jaw-dropping new polls since yesterday - maybe now that the first debate is over and there wasn't a clear Trump win they have uncertainty going down?

Nate said when they launched the model that if they told the model the election was that day, Biden's odds would be above 90%. So if Biden's numbers stay even, we should expect to see his odds increase as we get closer to Election Day.

Anyway, I made a simple swing/trend map from 2016 for the current forecast (at least, current as of a few minutes ago). Red is a Democratic swing and trend, gray is Democratic swing but Republican trend, and blue is Republican swing and trend. If this is accurate, then Biden's popular vote advantage is mostly coming from Republican states, not Democratic states. It's probably not very accurate, but it does make some sense given polls showing Biden doing reasonably well in rural areas.

Edit: fixed Vermont.

[The Mikado]

He said in the aritcle that went up last night that Biden's odds would be 91% on Election Day if things stay consistent.

Given that 91% was exactly the odds he gave Obama 2012 on Election Day when Obama only had a 1.5% national polling average lead over Romney, it says a lot about how much more cautious this model is that it'd give Biden's much larger lead the same chance.

[wbrocks67]

He said in the aritcle that went up last night that Biden's odds would be 91% on Election Day if things stay consistent.

Given that 91% was exactly the odds he gave Obama 2012 on Election Day when Obama only had a 1.5% national polling average lead over Romney, it says a lot about how much more cautious this model is that it'd give Biden's much larger lead the same chance.

At this rate, at least the Prez forecast seems to be converging on realistic results. The Senate forecast meanwhile is way too cautious and is still way too optimistic for most incumbents.

[Tintrlvr]

https://projects.fivethirtyeight.com/2020-election-forecast/

As of this morning, Biden now has a 79% chance.

Looks like it's up to 80 now, at least as I'm looking at it. No jaw-dropping new polls since yesterday - maybe now that the first debate is over and there wasn't a clear Trump win they have uncertainty going down?

Nate said when they launched the model that if they told the model the election was that day, Biden's odds would be above 90%. So if Biden's numbers stay even, we should expect to see his odds increase as we get closer to Election Day.

Anyway, I made a simple swing/trend map from 2016 for the current forecast (at least, current as of a few minutes ago). Red is a Democratic swing and trend, gray is Democratic swing but Republican trend, and blue is Republican swing and trend. If this is accurate, then Biden's popular vote advantage is mostly coming from Republican states, not Democratic states. It's probably not very accurate, but it does make some sense given polls showing Biden doing reasonably well in rural areas.

Edit: fixed Vermont.

More likely to be a problem with polling areas on the political extremes. At a certain point, you are statistically much likelier to have errors in favor of the minority position than in favor of the majority position. This shows up when polling demographic groups (Republicans always overpoll with black voters), but is also an issue when polling places like Wyoming (Democrats overpoll) or California (Republicans overpoll), at least as long as their actual results are very lopsided.

It’s not usually probed much because we mostly care about polling in close races.

[emailking]

He said in the aritcle that went up last night that Biden's odds would be 91% on Election Day if things stay consistent.

Given that 91% was exactly the odds he gave Obama 2012 on Election Day when Obama only had a 1.5% national polling average lead over Romney, it says a lot about how much more cautious this model is that it'd give Biden's much larger lead the same chance.

I'm not so sure about that. The model also detected structural advantages for the democrats in the electoral college in 2008 and 2012. But now it thinks Trump has a structural advantage.