seems to be bucking the trend. What is he up to?

Nate has Hillary at 86%. Which is very high! But if you look at other predictors, he's rather more conservative.

I think there are three issues:

First - which you can read about here - it looks like 538 uses a different statistical distribution model, known as a t-distribution, which has "fat tails." Basically it's quite conservative at saying that these odd events will not happen. There is a much higher statistical burden so to speak in moving from 85% to 95% than from 55% to 65%.

Second - he also seems (I may be wrong) to try to both normalize and make use of all kinds of polls by perhaps underweighing the individual levels in favor of making interferences from the intra-poll movement. So perhaps Rasmussen or LAT is a horrible poll; but if (unless the pollers are liars) one of these polls moves either to Trump or Hillary you can squeeze information from that, which he does.

Third - his predictions seem to be less stable and more prone to random statistical noise (e.g. which poll comes out when)

I think those three combine to make it unlikely he'll be at the heights of Princeton (at 99%) and also make it easy for us to see trends where there really are none, just random wiggles.

For what it's worth, if 538 is a touch more pro-Trump than others, I kind of appreciate it -- it's good not to believe our own crap. And the idea that Hillary has a 99% shot of winning is, to me, plainly nuts. 90% sure but the orange asshole could, somehow, unfortunately pull it off.

Lesson: donate, vote, and if it's your thing: Pray!

Nate is sort of OK on basic stats


But his collating various polls with various methodology is very very iffy

In fact. most polls are utter garbage (horrendous methodology and leading to non representative samples) and offer no useful data and adding or collating this data just means your contaminating your results.

If you have one garbage poll then we're on the same page. Agree.

But let's assume you've got a sort of crappy poll that comes out once a month that has a CONSISTENT methodology. (Half phone, half online, a certain weighing).

The month on month deltas can give you a low-to-medium indicator about the DIRECTION of the overall level.

Now, if the poll is kinda crappy AND the methodology is inconsistent, then it's just crap

Do you agree or do you think even the above is too generous?

People fixate to much on the stat, and not enough on how polls are done
Many polls are done to get talking points, for attention, and not to get true info

That was very clear in the avalanche of very very bad polls before the debate
Clinton was supposedly losing ground; well, I don't believe that one second.
The polls were closer than now, but she wasn't losing ground.

Getting a representative sample is what makes the stat actually mean something
MArgin of error and whatever mean nothing if you don't have that
You just don't know what the results are.

Using data from polls with the same methodology is OK if the methodology is sound and polls have a large enough sample, like 2048 for a national poll. For me, around 2000 is the minimal sample for a decent national poll which tries to get a representative sample.

For example of a bad poll method, you do a national poll of 500 people on land lines and call around supper time, use a terrible calling list which has more callers that are rural or in red state (rural people would tend to have more land lines in the first place),
use leading questions (or order the questions to bias the results, yes question order has an impact)

That poll would likely be absolutely terrible with a massive oversampling of GOP voters.

Also, a small sample means that any methodology biases will be amplified.

Then you add on the bias in the likely vs registered voter models and it just gets into extra garbage territory.

State polls can also be off if they're small enough and people are not careful what they do.
But states tend to have more consistent voting patterns which blunts the impacts of those problems if the poll is large enough.

It's a catastrophe.

Don't let some of our "friendly" poll-obsessives tell you different.

Troublesome for the GOP I'll bet.

is that their wild swings often put Trump ahead in a matter of days, and we are forced to say that Reuters suck.
Then, Clinton surges in a wild swing and we have no problem with Reuters.

Most polls have Clinton ahead by a large margin and do not swing wildly. I am simply ignoring Reuters.

last 3 daily additions ( its a 4 day survey period so they are now capturing 2 post debate days) has gone from Clinton +2 to +6 to +9. This trend should continue to grow as more post debate days are captured

I posted the 538.com site which groups last 3 Ipsos Polls


http://projects.fivethirtyeight.com/2016-election-forecast/updates/

Produces some interesting information:

Trump is losing with Hispanics by 43%

Trump is losing with Blacks by 70%

Trump is losing with Asians by 26%

Trump is losing with women by 10%

Trump is losing with 18-34 yr olds by 24%

These numbers mean there is no doubt this election is over.