|
Convergence doesn't mean they all become volatile. It's the opposite. They all stop assuming that a November surprise could happen, and so whatever forecasting they're doing gets mostly taken out of the equation because all the data that can be known from public polling about the election will be known. Basically, as we get closer to the election the possibility space for swings gets smaller to the point where after they stop putting polls in on election day it's very close to zero. So they probably won't swing very much unless the polls themselves swing a loving ton.
|
#
?
Nov 7, 2016 09:43
|
|
|
|
| # ? Jun 8, 2024 17:01 |
|
|
ErIog posted:I feel like candidates with record-setting unfavorables and one candidate having no GOTV to speak of makes the "this election is an outlier" defense a pretty good one. Nate only started using "well this election cycle is so different" during the primaries when, in defiance of all polls showing Trump ahead, he predicted that Rubio was the favorite long after he stood no chance of winning the GOP nom. Prior to that he was utterly confident of his "party decides" mantra to the point of snarkiness. Now in defiance of all polls showing Clinton ahead of even the margins Obama needed to decisively win re-election in 2012, his model is hyping up the possibility of a Trump win. I don't think it's impossible, but it should give anybody who has followed 538 over the past three election cycles pause.
|
|
#
?
Nov 7, 2016 09:43
|
|
|
they've basically already converged
|
|
#
?
Nov 7, 2016 09:45
|
|
|
ErIog posted:Convergence doesn't mean they all become volatile. It's the opposite. They all stop assuming that a November surprise could happen, and so whatever forecasting they're doing gets mostly taken out of the equation because all the data that can be known from public polling about the election will be known. Wouldn't the lack of forecasting mean they're more likely to swing in whichever direction the polls do? The now-cast is the one that has always assumed there's no time left, which is why it's been the most reactive one.
|
|
#
?
Nov 7, 2016 09:46
|
|
|
exquisite tea posted:Now in defiance of all polls showing Clinton ahead of even the margins Obama needed to decisively win re-election in 2012, his model is hyping up the possibility of a Trump win. I don't think it's impossible, but it should give anybody who has followed 538 over the past three election cycles pause. It seems pretty reasonable to me. Seems like he got lucky with an okay model in 2008. 2012 was pretty average as far as elections go. Republican wave in 2010 was a little crazy, 2014 was pretty average. 2016 is the first time it's really come up against a big challenge. I guess I've never considered Silver to be the complete amazing genius that I guess a lot of other people considered him to be. So his model's performance this cycle has struck me as completely expected. So it's hard for me to understand why people are having such an off the loving wall reaction to the ESPN data journalism's take on the US election.
|
|
#
?
Nov 7, 2016 09:47
|
|
|
Like I'm pretty sure if you knew nothing about the favorables/unfavorables of the candidates and wiped their names from the electoral map, it would look like a pretty much Generic R vs. D ballot with some electoral realignment among college-educated + non-college educated whites and women.
|
|
#
?
Nov 7, 2016 09:47
|
|
|
exquisite tea posted:Like I'm pretty sure if you knew nothing about the favorables/unfavorables of the candidates and wiped their names from the electoral map, it would look like a pretty much Generic R vs. D ballot with some electoral realignment among college-educated + non-college educated whites and women. I 100% agree with this.
|
|
#
?
Nov 7, 2016 09:49
|
|
|
exquisite tea posted:Like I'm pretty sure if you knew nothing about the favorables/unfavorables of the candidates and wiped their names from the electoral map, it would look like a pretty much Generic R vs. D ballot with some electoral realignment among college-educated + non-college educated whites and women. Trump is doing a lot better in the rust belt than a generic R, and a lost worse elsewhere (the South, Arizona, and other random states). Racial demographics are more important than ever for this election.
|
|
#
?
Nov 7, 2016 09:57
|
|
|
Supercar Gautier posted:Wouldn't the lack of forecasting mean they're more likely to swing in whichever direction the polls do? The now-cast is the one that has always assumed there's no time left, which is why it's been the most reactive one. It depends on the specific model, but when you stop accounting for time left you also remove acceleration from the equation. So if Hillary is up 2 points more in the aggregate tomorrow than she was yesterday, they're not going to assume that upward swing will continue on Wednesday because the election already happened. That's what makes some of the models very volatile when the candidates are within MoE or slightly outside the MoE in favour of the other candidate that was out of favor. They have a tendency to assume the race is swinging when it might not be. They're not going to do that on election day.
|
|
#
?
Nov 7, 2016 11:14
|
|
|
Vox Nihili posted:Trump is doing a lot better in the rust belt than a generic R, and a lost worse elsewhere (the South, Arizona, and other random states). Racial demographics are more important than ever for this election. I think it's certainly important for understanding the makeup of the electorate and why the red/blue lean of certain states have shifted in this cycle, but these are variables you can adjust for, and Nate's defense of "nobody could have foreseen this" is a little weak when we have an array of surveys showing Trump's unfavorables among Latino voters compared to Clinton with non-college educated whites. Compare the utter confidence he had in his language from the beginning of the primaries to now and it's even more stark.
|
|
#
?
Nov 7, 2016 11:51
|
|
|
Given most of these models are placing Clinton's Median EVs at 310 +/- 15, isn't the real difference between their win probabilities drive by their assumptions about uncertainty? You've got very low uncertainty (Huff Po and PEC) at 99%, Linzen, Upshot and others in the 80s and Nate at 70ish. Given the median is similar seems like the big difference is expected variance in results? Or have I forgotten my statistics.
|
|
#
?
Nov 7, 2016 13:22
|
|
|
Uncertain models regress to 50/50, so yeah
|
|
#
?
Nov 7, 2016 13:31
|
|
|
ErIog posted:The problem with Nate's model isn't anything to do with whether or not he uses a trend line adjustment or anything else in that stupid HuffPo article. Pretty much. He is becoming too dogmatic about data / empiricism and almost deathly afraid to actually use his judgment. And the idea that *any* poll tells you *something* is ... weird. Pedro De Heredia has issued a correction as of 13:49 on Nov 7, 2016 |
|
#
?
Nov 7, 2016 13:46
|
|
|
https://twitter.com/NateSilver538/status/795449708479217664 this is probably the most important tweet hes made on his model the last few days. If his goal is to reflect public polls, well okay, his model is great at that. I'm not sure why that is the goal, rather than using public polls as part of the data driven way to forecast actual election results. But if he is just trying to create a model of what public polls say rather than what will happen on on election day, he is doing a good job at that.
|
|
#
?
Nov 7, 2016 13:46
|
|
|
the one thing i'll say in nate's defense is its not good practice to adjust the model midway through, but he should definitely be rethinking it in the off-season (not that he will as he'll be shitcanned on wednesday by espn)
|
|
#
?
Nov 7, 2016 14:00
|
|
|
https://twitter.com/NateSilver538/status/795611533086691328
|
|
#
?
Nov 7, 2016 14:03
|
|
|
evilweasel posted:the one thing i'll say in nate's defense is its not good practice to adjust the model midway through, but he should definitely be rethinking it in the off-season (not that he will as he'll be shitcanned on wednesday by espn) You don't think ESPN is going to drive traffic with the CARMELO ratings?
|
|
#
?
Nov 7, 2016 14:08
|
|
|
evilweasel posted:the one thing i'll say in nate's defense is its not good practice to adjust the model midway through, but he should definitely be rethinking it in the off-season (not that he will as he'll be shitcanned on wednesday by espn) If ESPN fired their analysts for being wrong, none of them would have job.
|
|
#
?
Nov 7, 2016 14:12
|
|
|
Charlz Guybon posted:If ESPN fired their analysts for being wrong, none of them would have job. he won't be fired for being wrong, he'll be fired for being expensive and not worth it in the least because espn is in cost-cutting mode, the site is garbage for everything but politics and not very good at politics these days he knows it and it's ratcheting up the stress to 11
|
|
#
?
Nov 7, 2016 14:17
|
|
|
I'd assumed the main "problem" with Nate's model was that he's assuming very large uncertainties on the polling results on the principle that modern polls are garbage. This is sort of equivalent to saying that the model prediction is being distorted by garbage polls, although the extremely wide distribution for Clinton's predicted electoral votes is best explained by large estimated uncertainties in poll results. It's possible Nate will be vindicated to some extent after the election, as it's looking like hispanic turnout is much higher than expected in several states (which is another way of saying common turnout models are garbage and he was right to assume very large uncertainties). Does anyone know what uncertainty Nate's been assuming for national polling average? I've read ~5.5% (which is huge) but nothing from 538 itself. Taking into account correlations between state results is pretty defensible imo, if anything models that don't include that are missing an important feature.
|
|
#
?
Nov 7, 2016 14:18
|
|
|
evilweasel posted:he won't be fired for being wrong, he'll be fired for being expensive and not worth it in the least because espn is in cost-cutting mode, the site is garbage for everything but politics and not very good at politics these days I really like their sports coverage. 
|
|
#
?
Nov 7, 2016 14:20
|
|
|
Nocturtle posted:I'd assumed the main "problem" with Nate's model was that he's assuming very large uncertainties on the polling results on the principle that modern polls are garbage. This is sort of equivalent to saying that the model prediction is being distorted by garbage polls, although the extremely wide distribution for Clinton's predicted electoral votes is best explained by large estimated uncertainties in poll results. It's possible Nate will be vindicated to some extent after the election, as it's looking like hispanic turnout is much higher than expected in several states (which is another way of saying common turnout models are garbage and he was right to assume very large uncertainties). Does anyone know what uncertainty Nate's been assuming for national polling average? I've read ~5.5% (which is huge) but nothing from 538 itself. he clearly didn't think hard enough about minimum criteria for a poll to get in, because if he had a 73 person google surveys poll would not have made the cut
|
|
#
?
Nov 7, 2016 14:21
|
|
|
G-Hawk posted:https://twitter.com/NateSilver538/status/795449708479217664 I might not be reading his tweet correctly, but if he's saying his model is just an aggregator, that's totally disingenuous. Listen to the way they speak about the model on a 538 podcast. It's not supposed to be predictive? Really? They're talking about it like it's a scoreboard at a basketball game. Nate Silver, who made his career predicting the value of baseball players? http://www.baseballprospectus.com/article.php?articleid=25465 That Nate Silver? The guy who wrote a book about forecasting? He's not forecasting now? He's just an aggregator?
|
|
#
?
Nov 7, 2016 14:35
|
|
|
Lager posted:Oh, what's up fellow Bloomington-Normal goon? uncle Nick's and the bagel place are the only good things about Rockford and it makes me happy any time I see someone recognise them man growing up in that region sucked and I can't believe my high school friends went and got degrees then went back to Rockford for eternity
|
|
#
?
Nov 7, 2016 14:44
|
|
|
evilweasel posted:he won't be fired for being wrong, he'll be fired for being expensive and not worth it in the least because espn is in cost-cutting mode, the site is garbage for everything but politics and not very good at politics these days how expensive could 538 be honestly? I bet nobody under the top five make more than 35k
|
|
#
?
Nov 7, 2016 14:47
|
|
|
evilweasel posted:he clearly didn't think hard enough about minimum criteria for a poll to get in, because if he had a 73 person google surveys poll would not have made the cut Presumably Nate see's his "added value" as the ability to construct predictive models that can take into account all available polling data, along with realistic data-driven estimates of polling uncertainties. If you're just going to look at the most reputable polls and cut everything else then you don't need to do that much analysis, and can't justify paying a hotshot polling "genius".
|
|
#
?
Nov 7, 2016 14:48
|
|
|
theflyingexecutive posted:how expensive could 538 be honestly? I bet nobody under the top five make more than 35k they hired him away from NYT so I assume "fivethirtyeight" as an org has a hefty contract, most of which goes to nate
|
|
#
?
Nov 7, 2016 14:53
|
|
|
ErIog posted:That's why his "lackeys" in 538's editorial content say that Clinton is more likely to win than what the model projects. The number of dog poo poo polls being thrown into the model and new Hispanic voters being tossed out by most LV screens is something they constantly talk about. Actually, the biggest problem is the trend adjustment. It is certainly responsible for the biggest difference between Nate and others. When the race was stable at Hillary +6, his model gave her 88% chance while Linzer was 95%. If you don't trust me, go see his state by state thing by yourself. "Trend line adjustment" is responsible for a full 1.7% difference in estimated share of the vote in Florida. If you took away all the bad polls in his model, it would still not get the same impact on win probability as a 1.7% swing in aggregate numbers. His trend line adjustment is single handedly responsible for a 1.6 or 1.7 % swing in the vote in Trump's direction in every key swing state. Even the example that I used before: yes, including a 74 person sample poll from google messes up their Delaware predictions. But bigger than that one data point, the real difference is that their trend line adjustment gives Trump +1.5% over just poll aggregates. You can go down the list. The difference between pollster.com (which does filter out bad polls and which is what PEC uses) is most of the time just the trendline adjustment. Nate expects 48.1% to 48.1% share to the vote in Florida. Pollster gives Hillary a 1.7% edge there. Size of trendline adjustment that Nate uses? +1.4% for Trump, -0.3% for HIllary. Nate expects 48.1% Hillary, 48.1% Trump in NC. Pollster gives Clinton a 1.9% edge. Trendline adjustment: +1.3 for Trump, -0.3% for Hillary. Nevada, Nate expects 46.9% Hillary to 46.6% Trump. Pollster gives Clinton 2.1% edge. Trendline adjustment gives Trump 1.4% and Hillary -.3%. The inclusion of bad polls is bullshit, but that would be responsible for a couple of decimal points difference in most of the cases. Don't trust me, check for yourself.
|
|
#
?
Nov 7, 2016 15:02
|
|
|
evilweasel posted:they hired him away from NYT so I assume "fivethirtyeight" as an org has a hefty contract, most of which goes to nate So even if ESPN cuts it entirely, seems like 538 would just spin off doing it's own thing again at the least. They have way too much of a brand name built up now for it to just disappear. They'll be around next election, just with different people and not under ESPN's wing.
|
|
#
?
Nov 7, 2016 15:04
|
|
|
the inclusion of bad polls hurts the uncertainty more than the raw vote number, it's driving towards 50/50, not towards trump
|
|
#
?
Nov 7, 2016 15:03
|
|
|
paternity suitor posted:I might not be reading his tweet correctly, but if he's saying his model is just an aggregator, that's totally disingenuous. Listen to the way they speak about the model on a 538 podcast. It's not supposed to be predictive? Really? They're talking about it like it's a scoreboard at a basketball game. Right. If he wants to claim his entire role is just to try to figure out what the average of all available public polls is, okay. That isn't how he frames his site at all. And it isn't really a great way to try to forecast an election or likelihood of outcomes.
|
|
#
?
Nov 7, 2016 15:11
|
|
|
paternity suitor posted:I might not be reading his tweet correctly, but if he's saying his model is just an aggregator, that's totally disingenuous. Listen to the way they speak about the model on a 538 podcast. It's not supposed to be predictive? Really? They're talking about it like it's a scoreboard at a basketball game. Then, if they get pestered enough, they offer these half-responses to even the most milquetoast criticism or request for clarification around the conclusions they're drawing. No intention of continuing to follow 538 and its people after tomorrow. Grey Fox has issued a correction as of 15:20 on Nov 7, 2016 |
|
#
?
Nov 7, 2016 15:17
|
|
|
G-Hawk posted:Right. If he wants to claim his entire role is just to try to figure out what the average of all available public polls is, okay. That isn't how he frames his site at all. And it isn't really a great way to try to forecast an election or likelihood of outcomes. It's really hard for Silver to walk himself back on this now seeing as he wrote an entire book about how everything can be predicted by an accurate enough linear regression model and the mathematical soothsayers who can discern the signal from the noise.
|
|
#
?
Nov 7, 2016 15:30
|
|
|
theflyingexecutive posted:Nate silver is actively reinforcing the rigged narrative and the further away his model is from hrc's actual numbers (double digits, btw) the more likely there will be violence post-election hrc's actual numbers are not double digit
|
|
#
?
Nov 7, 2016 15:33
|
|
|
evilweasel posted:they hired him away from NYT so I assume "fivethirtyeight" as an org has a hefty contract, most of which goes to nate
|
|
#
?
Nov 7, 2016 15:34
|
|
|
rudatron posted:wow, seems like the nyt dodged a bullet Silver was a much better fit at NYT when he could just update once every couple days and wasn't prodded by the constant pressure to produce more clickbait content.
|
|
#
?
Nov 7, 2016 15:35
|
|
|
he probably bought into the libertarian-nerd fantasy of owning your own business and eventually becoming a millionaire, without realizing that being a statistician doesn't make you a good manager, and if any media organization has good managers, it'd have to be the nyt
|
|
#
?
Nov 7, 2016 15:38
|
|
|
mastershakeman posted:uncle Nick's and the bagel place are the only good things about Rockford and it makes me happy any time I see someone recognise them That's hilarious, I actually used to work at the bagel place back in the early 2000s. I can't imagine wanting to go back there after getting away for college. Glad to hear you made it out, though.
|
|
#
?
Nov 7, 2016 15:42
|
|
|
theflyingexecutive posted:the inclusion of bad polls hurts the uncertainty more than the raw vote number, it's driving towards 50/50, not towards trump No, it doesn't. Uncertainty created by bad polls is bad, but relatively small. A 1.6 to 1.7% swing towards Trump in every single swing state has a huge impact on overall probability of win. Nate actually shows his 80% confidence intervals for each state in his "who's ahead in each state and by how much." Having shorter tails would affect his model less than removing the trendline adjustment and moving all swing state results 1.6 to 1.7% to the left. When his model was giving Hillary a 1.7% expected edge in Florida, he gave her a 61% chance. That is certainly underconfident, but if right now his model gave out similar probabilities for Hillary in Florida, NC and Nevada, he likely would have Hillary's chance of winning at over 75%. As Sam Wang posted today, the only way his model can get close to Nate's level of uncertainty is if he assumed a+/-5% average polling error, which I don't think even Nate thinks is reasonable. So uncertainty in Nate's model is certainly a problem. But nowhere near as big as his trendline adjustment.
|
|
#
?
Nov 7, 2016 15:47
|
|
|
|
| # ? Jun 8, 2024 17:01 |
|
|
Can we get the United States Geological Survey to monitor the 538 offices for the next 48 hours? I need to see how bad the shaking gets.
|
|
#
?
Nov 7, 2016 16:37
|
|