I followed the 2016 election and 538 very closely. Over the last week or so before the election was called, 538 gave Clinton a 3 out of 5 (or 60%) chance of winning, dramatically lower than any other outlet anywhere, and 35% lower than the NYT prediction.
Anyone paying attention was pretty certain Trump was going to win that. All you had to do was look at Trumps subreddit vs Hilarys. Look at Trumps rally turnouts. The energy behind Trump was insane and it seemed pretty clear something was going on in the media to make it sound like only a "minority" of people supported him. That's why I bet around $1000 with some coworkers he would win that were certain he was going to get destroyed.
Note: This isn't me saying I support or don't support Trump.
> The energy behind Trump was insane and it seemed pretty clear something was going on in the media to make it sound like only a "minority" of people supported him.
Only a minority of people did.
What’s wild about 2016 is that people seem to have bought into the narrative that because it was such an upset of the mainstream media’s expectations, that somehow that translated into an actual landslide.
Trump didn’t win the popular vote, and his actual margin in the decisive states was razor thin as well.
Did he have more passionate fans than Hilary? Clearly. But a passion vote counts the exact same amount as a “lesser of two evils” vote. See: 2020.
Passion votes count the same, sure. But impassioned voters advocate for their candidates and get their friends to vote. Lesser of two evils voters don't really care like that.
It's the whole reason startups go for natural marketing. Users that naturally come upon and love your product will proselyte it more and better than anyone else.
I wasn't surprised Trump won either, but then I followed 538's coverage closely which made clear that he wouldn't win the popular vote, but could well win the presidency.
> Anyone paying attention was pretty certain Trump was going to win that. All you had to do was look at Trumps subreddit vs Hilarys. Look at Trumps rally turnouts.
By 2020, no one was underestimating Trump anymore, and it was clear people cared enough to vote him out for real. The energy at his rallies was no longer there. His subreddit, where the massive following was making strong waves during 2016 elections, was gone. Hillary not running anymore helped as well.
In 2020, obviously senile and charisma-void, candidate Joe Biden surprisingly manifested ~81 (Edit: actually 8) million more votes than the 'second-coming of Christ' Barack Obama before him, while speaking to low double-digit audiences at his rally stops. Really makes you think.
Edit: in 2020 Biden received 81,284,000 votes , Trump 74,221,000 votes, while in 2008 Obama got 69,498,516 votes
Thanks to the user remarking on the error. It's the kind of mistake that happens when you rewrite a comment ten times.
My apologies if I misunderstood your statement but I interpret it to mean that you're skeptical that Biden received a large number of votes compared to Obama.
If my interpretation is correct, then I think you're missing the big factor that conservatives and registered Republicans actively campaigned for Biden. They didn't just reject Trump, they explicitly said they were voting Biden, Campaigned for him, asked their fellow GOP members to vote Biden over Trump.
At the end of the day, you had Democrats + Independents + Conservatives voting Biden. That coalition didn't exist during Obama's time. In fact, the people who most likely voted for Romney were the people who were now voting for Biden
If your analysis is correct, you would expect Trump to have gotten a desicive victory in the popular vote. Not only did that not happen, but he lost the popular vote by 2 points.
Echoing a sibling comment, it seems strange it's so common that people look at probabilities like this and say that because the thing that happened was less likely that the model was wrong.
They say explicitly how this works: they run simulations and count winners. That means among their simulations is very likely to be an instance that is similar to the real outcome.
Yes, I agree. If you are tasked for example with analyzing a coin to assess its bias, you could do a very thorough job, measuring precisely its mass distribution, its contours, etc and come up with an extremely accurate estimate that say, the probability of it landing Heads is 73.22%.
If the coin is subsequently flipped and comes up Tails, it doesn't mean your analysis was wrong.
On the other hand, 538's simulation was based on probabilities of winning each state, and as I think most will agree, before and after the election, the most important states were Wisconsin, Pennsylvania, and Michigan. On election day, 538 gave the following probabilities:
Wisconsin 83.5% Clinton [1]
Pennsylvania 77.0% Clinton [2]
Michigan 78.9% Clinton [3]
I think Nate really screwed the pooch here. The vast bulk of his effort should have been spent on these states. If that is in fact the case, then I retract my criticism.
>That means among their simulations is very likely to be an instance that is similar to the real outcome.
So what? If you only make enough enough predictions then some of those will turn out predict the future correctly. It is how "The Simpsons did it first" meme came into being.
Depends, is the coin a Democratic party establishment candidate and the predictor part of the liberal politico-media complex trying to prime an audience?
> Trump is a significant underdog — he has a 13 percent chance of winning the election according to our polls-only model and a 23 percent chance according to polls-plus. But those probabilities aren’t that small. For comparison, you have a 17 percent chance of losing a “game” of Russian roulette.
A) Their more advanced model gave Trump a 23% chance, the 13% chance you mention is their simpler model.
B) Things with a 23%, 13%, and 7% chance happen all the time
C) You often have a better understanding of events after they happen relative to before.
If you want to show a predictor is unreliable, you need to look at a bunch of their predictions and see how they perform in aggregate. While also considering how correlated the results are (and, if applicable, predicted to be)
Their predictions were far higher than anybody else. I'm not sure what you're expecting them to do. They can't just ignore data because it "feels" like it's more likely that Trump will win. Its not a number they arbitrarily decide on. Based on the numbers they had, they made the most accurate estimation they could.
>Have you actually gauged 538’s own probabilities against themselves?
Yes, and the publish datasets of all their work to let others analyze them too. They're the most accurate of all the major sites.
Here's their github site [1]. It contains tons of datasets, past performance, analysis tools, everything you want to actually check if your claims are true or false, and most importantly, how accurate 538 has been over time.
Their seems to be a systemetic bias for them to be underconfident. That is to say, when they a result is likly (>50%) to happen, they underestimate how likely. Conversly, when they think an event is less likely (<50%), they overestimate its chances. Still, more likely events tended to occur more often than less likely events.
Correcting for this systematic error would mean they should have given Clinton an even higher chance.
While it should have been obvious to anyone who had been to Ohio in the couple years before the election that Trump was the favorite, 538 did give Clinton a lower chance than other outlets. Also, 13% things happen all the time. Now if the model had said 0.13%, ok, maybe you side-eye it
And the same 538 concurred with 87% chance of a Clinton win.
And the very same 538 that claimed a “Comey Effect” that purportedly botched their deep analytical effort, posthumously.
https://fivethirtyeight.com/features/what-a-clinton-landslid...