538 model & poll tracker thread

[The Mikado]

My new favorite. We can all blame HockeyDude for Trump for four years.

[wbrocks67]

It's miniscule but I really dont understand:

Before YouGov poll: Biden +8.3

YouGov poll which has +11 (RV) and +9 (LV) tabs

Biden's lead ticks *down* to 8.2.

How did it tick down when the poll result was *above* the previous average?

[Figueira]

It's miniscule but I really dont understand:

Before YouGov poll: Biden +8.3

YouGov poll which has +11 (RV) and +9 (LV) tabs

Biden's lead ticks *down* to 8.2.

How did it tick down when the poll result was *above* the previous average?

Could be that some other poll was removed from the average.

[dunceDude]

Warming to the model.

Unusual maps are expected.

[wbrocks67]

Does 538's forecast incorporate district polls?

I don't see how it can have Trump +5 right now if all of these district by district polls are being inputted as well...

[GeorgiaModerate]

Does 538's forecast incorporate district polls?

I don't see how it can have Trump +5 right now if all of these district by district polls are being inputted as well...

I doubt any of the models handle district-level polls, or would even want to.  If the district polls have implications to a statewide result (which after all is the important thing), then that should show up in statewide polls at some point.

[BaldEagle1991]

My new favorite. We can all blame HockeyDude for Trump for four years.

I'm gonna look at New Jersey and Nevada as I did Wisconsin and Michigan in 2016.

[Boobs]

Wow, Fivey Fox finally said something sensible!

[Non Swing Voter]

Trump wouldn't even need Nevada in that map if he flipped New Jersey.

[BaldEagle1991]

Wow, Fivey Fox finally said something sensible!

Fivey Fox is the weirdest mascot for FiveThirtyEight

[Mr. Morden]

Wow, Fivey Fox finally said something sensible!

Fivey Fox is the weirdest mascot for FiveThirtyEight

I constantly imagine Fivey speaking with the voice of the "Cluckin' Chicken":

[Fmr. Gov. NickG]

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?

[HagridOfTheDeep]

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.

[wbrocks67]

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.

[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.

[GeorgiaModerate]

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.

I think they specifically said that the time factor would start having more effect around mid to late September.

[Sorenroy]

TL;DR: Having outliers included in the state-by-state dot visualizations is confusing and unhelpful.


I've been wanting to do this for a while, but it's hard to catch the model when it's not updating and screwing up hours of work. I'm not a huge fan of the way the model is visualized overall, but there are a few key examples where the visualization is just downright unusable. Namely, the dots on each graph for the states have outliers so extreme that they just confuse the actual data. (I'm stealing demosboccaccio screenshot from the weird 538 map thread here.)



As you can see, the model indicates that Biden has a >1% chance at winning West Virginia, but, for whatever reason, the dot graph includes an outcome where Biden still wins the state. This is the case for every single state. The most extreme of these is Nebraska's third congressional district, where Biden has a 0.03% chance of winning yet there is still an example shown where Biden wins the district. If the point of the dot scatter plot/graph/whatever is to showcase a good range of possible samples, as is indicated by the description, then why include such outlandish outliers? The chance of winning and popular vote graphs below are much better at showing off the competitiveness of a state, but they are below the fold and require scrolling while those dots are put front and center. Just to emphasize this, I have included five maps made from the given data. The first three use true model outputs: the current map and maps of all states Biden and Trump have at least a 0.5% chance of winning. Finally, I went through state by state and made a map showing the race using the most extreme dot from every graph.

Current Model Output:

Joe Biden — 319 EV (52.7%)
Donald Trump — 219 EV (46.0%)

0.5% Chance for Biden:

Joe Biden — 534 EV
Donald Trump — 4 EV

0.5% Chance for Trump:

Donald Trump — 438 EV
Joe Biden — 100 EV

Biden's Best Dots:

Joe Biden — 538 EV (64.4%)
Donald Trump — 0 EV (34.0%)

Trump's Best Dots:

Donald Trump — 535 EV (58.5%)
Joe Biden — 3 EV (40.3%)

Obviously, the 0.5% chance maps are completely unrealistic, but at least they're based on some aspect of the model (and come with the caveat that they're 1/200). The whole point of running the model 40,000 times is to make it so you don't need to rely on any specific output but can, instead, rely on the average. By including such outlandish outcomes, you end up getting something that does not show anything at all.

[Smash255]

It's miniscule but I really dont understand:

Before YouGov poll: Biden +8.3

YouGov poll which has +11 (RV) and +9 (LV) tabs

Biden's lead ticks *down* to 8.2.

How did it tick down when the poll result was *above* the previous average?

They also extrapolate state poll data into their national #'s, so if there was some state poll data showing better results for Trump released around the same time  it might have canceled out the increase from a good national poll.  Older national polls or even older state polls that were strong for Biden receiving less weight could have been a factor as well.  They are now at 9.2 as of this morning

[Fmr. Gov. NickG]

It’s really bothersome how little information the design of this model makes easily available to viewers.  It seems like the only message they want to communicate is “I guess anything could happen?!?”.

Where’s the “polls only” forecast?
Where’s the “now cast”?
Where’s the easy link to the database of polls used to generate the projections?
Where’s the accounting for how individual “fundamentals” factors influence the projections?
Where’s the map??

All of this was available in the 2016 model, which actually performed quite well.  This is such a huge step backward.

[The Mikado]

It’s really bothersome how little information the design of this model makes easily available to viewers.  It seems like the only message they want to communicate is “I guess anything could happen?!?”.

Where’s the “polls only” forecast?
Where’s the “now cast”?
Where’s the easy link to the database of polls used to generate the projections?
Where’s the accounting for how individual “fundamentals” factors influence the projections?
Where’s the map??

All of this was available in the 2016 model, which actually performed quite well.  This is such a huge step backward.

He has said since 2017 that he's never doing another Nowcast and it was a mistake to make it, so I'm not surprised it's gone.

The rest, though, should be there. I really want that list of probabilities of various interesting things (Electoral tie, landslide, EV/PV split, etc) that they had in 2016.

[Devout Centrist]

It’s really bothersome how little information the design of this model makes easily available to viewers.  It seems like the only message they want to communicate is “I guess anything could happen?!?”.

Where’s the “polls only” forecast?
Where’s the “now cast”?
Where’s the easy link to the database of polls used to generate the projections?
Where’s the accounting for how individual “fundamentals” factors influence the projections?
Where’s the map??

All of this was available in the 2016 model, which actually performed quite well.  This is such a huge step backward.

But, uh, we need more uncertainty in the models! We need a section for New York Times headlines! And, uh, Fivey Fox!!!

[Sorenroy]

It’s really bothersome how little information the design of this model makes easily available to viewers.  It seems like the only message they want to communicate is “I guess anything could happen?!?”.

Where’s the “polls only” forecast?
Where’s the “now cast”?
Where’s the easy link to the database of polls used to generate the projections?
Where’s the accounting for how individual “fundamentals” factors influence the projections?
Where’s the map??

All of this was available in the 2016 model, which actually performed quite well.  This is such a huge step backward.

He has said since 2017 that he's never doing another Nowcast and it was a mistake to make it, so I'm not surprised it's gone.

The rest, though, should be there. I really want that list of probabilities of various interesting things (Electoral tie, landslide, EV/PV split, etc) that they had in 2016.

There is a link at the very bottom of the "National overview" page that will download model outputs. Besides more precise numbers, there is a file called "Scenario Analysis" that includes the chances of:

1: An Electoral College tie. (0.3%)
2: A presidential recount. (0.5%)
3: Trump winning the popular vote. (16.6%)
4: Biden winning the popular vote. (83.4%)
5: Trump winning the popular vote but losing the Electoral College. (0.1%)
6: Biden winning the popular vote but losing the Electoral College. (10.2%)
7: Trump winning a majority. (12.3%)
8: Biden winning a majority. (78.5%)
9: Trump winning by over ten points. (1.2%)
10: Biden winning by over ten points. (31.9%)
11: The same map as 2016. (0.0%)
12: Trump winning a Clinton state. (41.1%)
13: Biden winning a Trump state. (88.0%)

All of this is interesting, and it's valuable to have, but it is definitely odd that it requires a separate file download to find.

[The Mikado]

It’s really bothersome how little information the design of this model makes easily available to viewers.  It seems like the only message they want to communicate is “I guess anything could happen?!?”.

Where’s the “polls only” forecast?
Where’s the “now cast”?
Where’s the easy link to the database of polls used to generate the projections?
Where’s the accounting for how individual “fundamentals” factors influence the projections?
Where’s the map??

All of this was available in the 2016 model, which actually performed quite well.  This is such a huge step backward.

He has said since 2017 that he's never doing another Nowcast and it was a mistake to make it, so I'm not surprised it's gone.

The rest, though, should be there. I really want that list of probabilities of various interesting things (Electoral tie, landslide, EV/PV split, etc) that they had in 2016.

There is a link at the very bottom of the "National overview" page that will download model outputs. Besides more precise numbers, there is a file called "Scenario Analysis" that includes the chances of:

1: An Electoral College tie. (0.3%)
2: A presidential recount. (0.5%)
3: Trump winning the popular vote. (16.6%)
4: Biden winning the popular vote. (83.4%)
5: Trump winning the popular vote but losing the Electoral College. (0.1%)
6: Biden winning the popular vote but losing the Electoral College. (10.2%)
7: Trump winning a majority. (12.3%)
8: Biden winning a majority. (78.5%)
9: Trump winning by over ten points. (1.2%)
10: Biden winning by over ten points. (31.9%)
11: The same map as 2016. (0.0%)
12: Trump winning a Clinton state. (41.1%)
13: Biden winning a Trump state. (88.0%)

All of this is interesting, and it's valuable to have, but it is definitely odd that it requires a separate file download to find.

This is very cool. Also lol re: 0.0% same map as 2016.

Trump by over 10 points at even 1.2% is stunning to me. If he's saying Trump has a 12.3% chance of even winning the popular vote at all, that "Trump wins by 10" makes up a tenth of those scenarios is shocking.

[Beefalow and the Consumer]

I imagine some of these outlandish scenarios require some sort of... extreme demographic event.

[emailking]

I imagine most of it is the model accounting for the possibility that the polls are all way off because of the same systematic errors. (But it applies this to both Biden and Trump which is why some huge Biden wins are possible.)