You misunderstand what Wang has said.
He said, from extremely early on, that the polls were showing Clinton up by 4ish and that it would be highly likely for the polls to travel within +/- 3 of that and to end +/- 3 of that. In fact the playout of this election has proven him to be correct. The worst for Clinton best for Trump news cycles and periods never pushed it beyond that 4 - 3 for more than the briefest of times and the best for Clinton worst for Trump rarely pushed it beyond that 4 + 3. Moreover as soon as the pressure from the news cycle abated back towards 4 it went each time.
The signal has been a consistent C+4 even though through it there has been a significant amount of noise. His model is relatively noise insensitive at the risk of possibly filtering out true signal; Silver has been the most noise sensitive, taking ever pop of static as potentially significant.
If accurate, obviously that is independently a blow to the Trump campaign and nationally not a terrific sign for him.
Trump CAN win the election without Nevada, if the demographics are mixed up a bit, which they could be. An Angry White Guy wave could deliver Pennsylvania, for instance, but he’s still got to take must win states like Florida and NC.
ETA: Rolston’s analysis is supported by other sources as well. It seems very likely Nevada is going Clinton. This has significant implications if true. Obviously it’s six EVs that Trump really could have used, but if it’s true Latino voters are unusually motivated to go to the polls, that could represent a significant polling error that moves the needle towards Clinton. Is this a Latino turnout increase unique to Nevada? If so, Trump has to make up the six EVs somewhere else, which he could do. It’s it’s a general trend for Latino voters to turn out and vote Democratic, though, it makes a difference in Trump’s chances in Florida, without with he cannot win, and could even mean Arizona is more of a tossup that we have been assuming.
As to scoring after the fact - asked here in this very thread. Click but I agains ask if the Brier score is best to use.
I’d be very interested in seeing this discussion continue elsewhere, because I’m suddenly a bit confused about this. Has anyone started a new thread for it already?
During this election, he’s shown a wide range of values for Trump’s chance. Which do you compare with the result?
Actually, I think you can run a check on the model even ignoring the final result! For example, if a probability is shown at 20%, how often does it later rise to 40%? This should happen just half the time!
If a “fair game” is one with zero vigorish, you have exactly zero expected return for any strategy.
Surely the amount by which the forecast probability changes over time may vary widely for different types of event & model? There’s no reason it need change at all for an event where the parameters are known with a high degree of certainty at an early stage.
But it’s not “noise” and it’s nonsense to suggest that it is. The movement in the polls is reflective of volatile movement in an election with a large number of uncommitted voters. There is a lot of uncertainty in this election due to the fact that these two candidates are extremely unpopular, something which resulted in third party candidates that, while not in any way competitive, are still not insignificant. If Trump weren’t capable of actually winning the race, we could call it “noise” but when the movement in polls, if fortuitously timed, could actually result in a candidate winning the electoral college vote despite losing the popular vote, I don’t think “noise” is the appropriate word.
I would be careful about writing off Trump in Nevada just yet. The odds continue favor Clinton as they have throughout the race, but we could still see one last surge on election day. For the record, I’m not betting against Clinton and I agree she’s the odds-on favorite, but there’s not enough good data to make a call.
Seems like the easiest way to reconcile Wang and Silver is to posit that the chance of a popular vote victory for Hillary/electoral vote victory for President-Elect Trump is even higher than the ~10% Nate is currently giving it.
Right now I’ve got Hillary winning with anywhere from 278 to 307 electoral college votes on election night. I suspect that she and Trump will probably split the big four of OH, NC, PA, and FL. The one thing I’m possibly underestimating is Hillary Clinton’s ground game advantage. I think it could very well help her get at least a split between Florida and North Carolina, but who’s to say it couldn’t get her both states and possibly Ohio as well? I’ve not seen enough good polling data to convince me that she could get much above 300 EVs but Jim Messina seems to think it’s entirely possible and he’s got a pretty good track record when it comes to delegate math.
The probability will finish at 0 or 1. If you assume it’s an accurate model and varies continuously, it is a trivial mathematical theorem that from a point N (N < 0.5) it has precisely a 50% probability of getting to 2N. (Or more generally a probability p of getting to N/p when N < p < 1.)
ETA: Yes, strange things can happen. But those strangenesses and their chances are factored into an accurate model.
I don’t think the assumption of continuity is a valid approximation. The situation here is that you are using evidence to update estimates of parameter values of a probability distribution, with the outcome modeled as a random variable. Now, obviously the random variable part in politics does not represent true randomness, it represents the sum of unknowns. In a situation where unknown events that are initially modeled as randomness are all revealed gradually over a long period of time, then I think continuity holds as an approximation. But not here - essentially we end up with a final probability on the morning of the election, then it jumps to 1 or 0 very fast. In order for continuity to be valid as an approximation, Silver would have to continuously update his probabilities on election day during that rapid move from 0.35/0.65 to 0/1 or 1/0, based on election-day information that is different in form from the prior information, and that updates very rapidly. I don’t think it’s reasonable to expect his model to do that, because real-time updating on election day iss just a very different task from predicting a week or even a day out.
How little movement in the polls do you need to not call it “volatile”? Since the conventions the aggregate Clinton lead has been between 1 and 7. If you look at 538’s own estimate of popular vote percentages, the curves are pretty flat
This HuffPo article (yeah, I know) says that the reason Nate Silver’s prediction is such an outlier is because he’s “unskewing” the polls in the name of “trend line adjustments”.
That … certainly sounds like unskewing. But don’t all the pollsters use trendlines?
I sort of thought that Sam Wang’s 90-something prediction is based partly on looking at what the polls have been doing for the last few months. Clinton’s been leading all this time, meaning that it’s more likely that she will continue to lead, barring any sudden alien abductions.
The article continues:
That seems odd but I don’t know enough to really know.
Silver responded on Twitter, saying, that they make national trendline assumptions because “empirically, that’s what works”.
Which sounds like post hoc reasoning, if you ask me, but maybe that’s Twitter’s fault.
It is weird that Silver is so out of line with everyone. If he had one close to his analysis, I’d trust it more. It looks to me like he’s trying to hedge his bets, saying that Clinton will win but fiddling his numbers so that if Trump wins, he can say - no look, I was in my margin of error, I still win!
I can’t help but suspect that Silver maybe has some contractual bonuses riding on calling this election correctly, so he’s doing everything he can to have it both ways.
But that’s probably my statistical ignorance talking.
At any rate, the more I read, the more I think skipping the national polls makes sense, since the national vote is unimportant.
Is a swing of 2 or 3 points meaningful? Or is it noise? One point Wang makes is that a race with such little variance is remarkably stable, historically. Not volatile. Historically, races commonly had swings of 30 or 40 points. Nothing like that has occurred this year.
But we don’t need 30 or 40 point swings to determine the election, do we? Maybe the term “volatile” requires an operational definition. I’m not in a position to define it, only to point out that the kinds of swings that have occurred are not insignificant and, depending on the time, could determine the outcome of an election. The presidential election is not one race, but 51 individual contests on the same day, the sum of which is then added to produce a final national result.
Of course nobody is seriously submitting that a single news event, in and of itself, is going to cause a 10-15 drop in the polls. But a series of events, such as the Trump’s leaked tax returns, his “Grab 'em by the pssy" comment, and the multiple victims who came forward were not just noise - they shed serious light on the character of a candidate and caused some to stop taking his candidacy seriously. On the reverse side of that coin, the problems with Obamacare, and the specter of more legal and post-election political problems for Hillary Clinton have dampened enthusiasm among her voters. It’s the difference between a voter saying “Yes, I’m willing to drive across town and stand in line for an hour or two” and saying "I don’t really like Trump but fck it, Clinton isn’t really going to make much of a difference in my life either.”
You’re the one who used the word “volatile” so yeah maybe you should have a working definition. Wang thinks polls staying Clinton +4 +/-3 is stable.
If Nate Silver had hard data to fall back on, I’m sure he’d make a bolder prediction like he did in 2012 when he said that Obama had an extremely high chance to win when polls were merely giving him a 1 point advantage. What Silver is saying is that the race now is fundamentally different, and I think he’s 100 percent right. With Gary Johnson and Jill Stein collectively pulling in probably more than 5 percent in the polls, this is the strongest third party factor we’ve had since Ross Perot in the early to mid 1990s. Nationally, those numbers seem small. In races like NH, FL, NC, NV - all of which could conceivably come down to 1-2 percent - that’s 50 electoral votes potentially impacted. Throw in undecided voters and it becomes even more complicated. A 2 to 4 point swing in the polls isn’t noise. It’s the difference between being president-elect and giving a concession speech.
Instead of refreshing 538, please consider making some GOTV calls today: