Thank you, New York Times.
They have a handy chart which lists the state-by-state probabilities for six different poll aggregators, and with a bit of tweaking on my side I’ve been able to turn it into a spreadsheet that tracks how well each prediction held. I’ll try to update it as results roll in, but even if I lag behind you can always download a copy for yourself and fill in the blanks.
While you’re there, you can have some fun. Two tabs are devoted to testing what would happen if Five Thirty Eight got their probabilities right, verses the Princeton Election ‘s predictions. A third is a “fantasy” one, which you can fill in with whatever finish you want and see who comes out on top.
Now, the New York Times seems to be a bit behind on their assigned probabilities; Five Thirty Eight’s moved on to give a better shot of Clinton winning Florida, for instance, and I noticed Princeton’s probabilities don’t line up either. So I’ll also try to add another column for both, updated with the last numbers from each.
Technology permitting, I also plan on doing a live blog of the results as they roll in, which will play out below the fold. It’ll provide a handy archive as the evening clicks past.
2:15PM EST: There’s no data yet, and there won’t be for a while: the earliest polls no sooner than 6 PM EST. That coincides with a time I’ll be busy, alas, so don’t expect instant updates. In the meantime, don’t put too much stock in exit polls, they’ve been wrong in years past. I also managed to put a coat of new paint on the spreadsheet, and inserted the state-by-state numbers from my run of Princeton’s election code.
In other news, the first voting lawsuit has been filed; high Latino turnout in Nevada kept the polls open late, and Trump’s pissed off about it. The suit sounds meritless, since Nevada law says “closing time” means “stop accepting new voters” and not “stop voting,” but I’m no lawyer nor legal expert.
3:12PM EST: The lawsuit itself. It says more about what’s in the head of Trump’s team than about any potential illegality:
This action is to preserve the status quo so that, if the election of presidential electors from the State of Nevada is contested, the candidates and courts must redress egregious violations of Nevada election law perpetrated by the Registrar on November 4, 2016.
4:23PM EST: Well, that was quick. The judge rejected Trump’s lawsuit, because A) keeping post-close ballots seperate is already a legal requirement, B) publicizing the names of the poll workers would almost certainly lead to intimidation, and C) they filed without notifying the state, as per procedure. An embarassing display of ineptitude.
I should also point out I’m not the only one evaluating the prediction markets. Buzzfeed’s approach also brought up something called a “Brier score,” which is commonly used to grade prediction accuracy. I’m not a fan of them, I think the likelihood interpretation is more intuitive and makes it easier to do a statistical analysis. Still, it’s all just numbers on a spreadsheet, so I added a page that tracks Brier scores too.
Since there isn’t much action yet, I’m filling in the “Fantasy” tab with the results from VoteCastr. It tries to use demographics and turnout information to guess how people are voting. Nothing’s been tried like this before, so it’s in the appropriate tab.
Speaking of which, did I forget to link to the spreadsheet in my original post? Whoops, consider that fixed.
5:40PM EST: Those Philadelphia complaints look like hot air, too. I read talk of extending North Carolina’s voting hours, due to glitches, and apparently Colorado’s voter database crashed for half an hour but is back up.
Exit polls seem to be released but not public. There’s no real surprises.
7:05PM EST: Damn, there was a shooting near a polling station in California. One confirmed fatality, but the news is still unfolding.
7:15PM EST: Ooo, some first results! Trump takes Kentucky and Indiana, Clinton takes Vermont. The “Fantasy” tab is favoring the New York Times aggregator, and prefers Princeton’s model to 538’s; there isn’t enough info for the actual results to be meaningful, at least yet.
8:15PM EST: Results are flying in now, with Clinton outperforming her polls and taking an early lead in the Electoral College vote by 68-37. Since these were all in states that the poll aggregators agreed on, though, the needle is barely moving.
I’ll have to go dark for a bit, alas, so don’t expect many updates for a bit.
11:20PM EST: I’m back! What did I miss?
…. oh. Oh dear. Unfortunately, Trump has started outperforming his polls, and states that used to be a Democratic lock like Wisconsin and Michigan now have slight Republican leads. The Five Thirty Eight model suggested a volitile race, and that’s coming to pass. The betting markets now favor Trump, with Election Betting Odds giving him a 90% chance as I type.
Ironically, that’s not showing up in my spreadsheet Close races tend to finish counting later, so states with a clear lead dominate the early results. This penalizes models which are fuzzy on the odds, like Five Thirty Eight’s, so even as its predictions come true the hard evidence for that will be slow to arrive. The Huffington Post’s aggregator has been at the top of the heap for some time, and Princeton’s model is beating Five Thirty Eight’s.
12:20AM EST: Looking South, I see a definite gloom. The Latino vote spiked early, but exit polls have since suggested they’ve faded. African-americans were never there. American friends and acquaintences are sharing black humour memes. Pennsylvania and Wisconsin are looking more and more remote.
But the Princeton Election Consortium’s model is the current front-runner. I was increduluous at that, and so I double-checked my entries. It’s legit. In fact, it’s still in line with what I said an hour ago. When I transferred the results to the “Fantasy” tab and flipped Wisconsin and Pennsylvania to Trump, Five Thirty Eight jumped ahead. The spreadsheet is just one or two surprise results from changing around. Either Clinton pulls off a miracle, or Princeton’s model is due for a dive.
November 9th: I’m still updating that spreadsheet, by the way. A number of states are still being counted, but it looks like the Princeton model will cling to its lead until the very end. Once all states are counted, I’ll write up an analysis.