538 model & poll tracker thread

[Wisconsin+17]

After getting his ass whupped last election, Nate is going with Biden having a 90 percent chance of winning?

Wow. Crazy sauce.

[Wisconsin+17]

According to our forecast, Pennsylvania is the most likely tipping-point state, and a lot of Biden’s chances in the Electoral College hinge on what happens in Pennsylvania. He leads Trump there by about 5 points in our polling average, but it’s not as large a margin as Biden might like. Last week, we gamed out what would happen if Biden lost Pennsylvania but won other Midwestern states like Wisconsin. (TL;DR there’s no clear Plan B for Biden.)

it’s pretty unlikely that either of them will hit the 270 electoral votes needed to win

That doesn't sound like 90 percent to me. The problem is Silver's model.

[Wisconsin+17]

Nate actually didn't mark Clinton as a sure winner in 2016 and he never said Clinton had it in the bag. In 2020, he's cautioned that Trump can still win multiple times.

It's obvious that there was a significant problem with his model in 2016. I would not be surprised by a 60 percent prediction for Biden in this election. That would seem reasonable and prudent to me.

But 90 percent against incumbancy advantage? Yeah, no.

Can you get anyone to put up a 10:1 bet against Trump, where a Trump election victory gets you 10x what you put down?

[Wisconsin+17]

Nate was among the most bullish forecasters on Trump.

The problem is he was still wrong. It doesn't matter that he was less wrong than others who were further out - the problem is that his model was significantly wrong.

Now, there's lots of reasons for it (mostly due to GIGO effects involving bad polls), but given that his model failed last time around, would behoove Silver to correct those problems so that the next time around, his prediction would be more accurate.

His model had Clinton at 71% on Election Day and in the 60's a few days earlier.  A 30% chance (Trump's in 2016) is nothing to sneer at; it's the chance of a .300 hitter in baseball getting a hit in a particular at-bat.

Yep, he was still wrong though.

(And it should be remembered that even 10% outcomes happen sometimes.  Ever had someone ask you to pick a number from 1 to 10, and you got it right?)

Problem is that his model is demonstrating significant evidence of systematic bias. If you're running a machine that consistently veers one way, and only one way, that means that you've got a problem with the protocol that you are using. Until you go under the hood and fix the protocol, you will still veer to the same direction.

His own statement seems to indicate this is a 55-45 Biden election (PA is the only route for Biden), so if Biden is up in PA, then that's the only path for him. Trump has similar issues with Ohio + Florida.

You could even conclude that despite being down in the battlegrounds, that Trump has an advantage due to having multiple paths wrt wisconsin, Michigan, Pennsylvania and Minnesota.  

[Wisconsin+17]

People just show how little they actually know about probability and models when they say this stuff.

So what are his error bars? How many standard deviations is he out? What happens if we feed different numbers into the machine - ie, assume a neutral electorate? What are his assumptions regarding states and their margins?

[Wisconsin+17]

For Florida he is assuming that Biden is +2.4 here.

Since Florida is an R+4 state, Nate is assuming a +6.4 Biden electorate nationally.

A D+2 electorate produces a Trump victory in Nate's model.

He's also assuming that Michigan, and Wisconsin are less favorable for Trump than Nevada (which seems wrong).

[Wisconsin+17]

But Trump's chances go up to almost 70% if you give him PA.
If you give him Wisconsin it goes up even more.

The problem seems to be the data Nate is feeding in the machine, specifically the numbers for Michigan and Wisconsin. Also, he only has Texas going 1.4% in favour of Trump.

Ohio is also only .4 for Trump. Yeah, Nate's staked his reputation on this election here. He's going to regret calling it a 90 percent election for Biden.

[Wisconsin+17]

Everything about this is wrong. Just because Clinton was the favorite doesn't mean he was wrong. It means that things Silver expected to happen about 30% of the time happened. Also, where's your evidence that his model consistently overestimates Democrats?

There's a couple of things that could be wrong.

One, there is the statistical model. How you test that would be to input different paramaters (national polls, outcomes for the state), etc. If you put the parameters in for the election and it gives you numbers that are far out from what occurred, there's something wrong in how you are estimating probabilities.

Two, if you input the election results and it gives you results that are close to the actual result, then you can be relatively sure that the model itself is working, but that the problem are the inputs - GIGO.

Has Trump ever even been in striking distance in Minnesota?

He was last election. Trafalgar has had him close.

I can't find a poll that has him close, and most polls have Biden's lead in the double digits.

Again, GIGO. Polls for 2016 did not have Trump within a point in Minnesota, hence the polls would not accurately assess Trump's chances of winning Minnesota in 2016.

Frankly, it seems increasingly clear that Wisconsin and Michigan are gone for Trump too. So, in actuality, it seems like your logic is bullish for Biden. Trump's most plausible victory scenario is 2016 - MI and WI, whereas Biden might not even need PA (which he's favored in anyway) because he's also favored in states like NC, GA, FL, and AZ and close to winning states like OH and TX.

See above.

There hasn't been an election since Carter in 1980 when an incumbent has been down 6.5 nationally. Reagan himself was only up 9.75.

The only two are Carter in 1980 and Hoover in 1932.

Are you really saying that Biden is going to lead the next transformative election on par with 1980 and 1932? Really?

Because that's what Nate is saying.