[HN Gopher] An election forecast that's 50-50 is not "giving up"
       ___________________________________________________________________
        
       An election forecast that's 50-50 is not "giving up"
        
       Author : luu
       Score  : 27 points
       Date   : 2025-03-08 22:20 UTC (2 days ago)
        
 (HTM) web link (statmodeling.stat.columbia.edu)
 (TXT) w3m dump (statmodeling.stat.columbia.edu)
        
       | bryanlarsen wrote:
       | Nate Silver also had an article saying that the most likely
       | scenario would be a blowout. There was a 25% chance that Trump
       | would win all 7 battleground state, a 15% chance Harris would win
       | all 7. No other permutation of the 7 states was anywhere close.
        
         | bombcar wrote:
         | If they're not completely independent variables, that should be
         | the expected outcome (e.g., they all trend the same direction).
         | 
         | And if they're not that close, they're not battleground.
        
           | fullshark wrote:
           | They are where the campaign money and attention is going,
           | what about battleground implies independence? Whoever wins
           | them will win the election.
        
         | chicagobob wrote:
         | True, but the word "blowout" in this case is just a crazy side-
         | effect of our weird electoral college system.
         | 
         | Everyone knows that in all the swing states (except Arizona),
         | the final vote margin was just a few percent, and that was well
         | within the MOE for all the "50-50" polling in each of those
         | states.
         | 
         | No one seriously believes that any President has had a blowout
         | election since maybe Obama in 2008 or Bush in 2004, but the
         | media sure loves the word "blowout".
        
           | nightski wrote:
           | So basically, if you ignore how the entire system works then
           | it wasn't a blowout lol. I'm guessing the media was taking
           | into account that we indeed use an electoral college system
           | so that is all that matters.
        
         | culi wrote:
         | Trump won the popular vote by 1.5%. That's the 8th closest
         | election in all of US history.
         | 
         | Maybe he meant an EC blowout. But that's easier to predict.
         | Most polling had almost all the swing states as extremely
         | close. The outcome was likely to swing in the same direction
         | across all swing states so an EC blowout is likely
        
           | wakawaka28 wrote:
           | As I recall none of the big polls the mainstream media was
           | pushing projected Trump to win. They even refused to call the
           | election until the wee hours of the morning, when the results
           | were pretty clear. Just like in 2016, the polls were intended
           | to deceive people into thinking there was no hope for an
           | alternative candidate.
        
       | bitshiftfaced wrote:
       | An accurate model can output a 50-50 prediction. Sure, no problem
       | there. But there is a human bias that does tend to make 50% more
       | likely in these cases. It is the maximum percentage you can
       | assign to the uncomfortable possibility without it being higher
       | than the comfortable possibility.
       | 
       | 538 systematically magnified this kind of bias when they decided
       | to rate polls, not based on their absolute error, but based on
       | how close their bias was relative to other polls'
       | biases.(https://x.com/andrei__roman/status/1854328028480115144)
       | This down-weighted pollsters like Atlas Intel who would've
       | otherwise improved 538's forecast.
        
         | culi wrote:
         | I'm not sure how to verify your comment since 538 was cut by
         | ABC a month or 2 ago. But Nate Silver's pollster rating
         | methodology is pretty much the same as 538's was during his
         | tenure there and can be found here:
         | https://www.natesilver.net/p/pollster-ratings-silver-bulleti...
         | 
         | It actually explicitly looks for statistical evidence of
         | "herding" (e.g. not publishing poll results that might go
         | against the grain) and penalized those pollsters.
         | 
         | In both rating systems, polls that had a long history of going
         | against the grain and being correct, like Ann Seltzer's Iowa
         | poll, were weight very heavily. Seltzer went heavily against
         | the grain 3 elections in a row and was almost completely
         | correct the first 2 times. This year she was off by a massive
         | margin (ultimately costing her her career). Polls that go
         | heavily against the grain but DON'T have a polling history
         | simply aren't weighted heavily in general.
        
           | tantalor wrote:
           | 538 was cut by ABC 6 days ago.
           | 
           | https://archive.is/E2nre
        
             | culi wrote:
             | Wow thanks. As a former regular reader it's felt a lot
             | longer
             | 
             | They've had several major cuts in the past couple of years
             | so maybe that's why it's felt like that
        
           | bitshiftfaced wrote:
           | > I'm not sure how to verify your comment
           | 
           | Here's how 538 explains how they factor in bias into their
           | grade:
           | 
           | > Think about this another way. If most polls in a race
           | overestimate the Democratic candidate by 10 points in a given
           | election, but Pollster C's surveys overestimate Republicans
           | by 5, there may be something off about the way Pollster C
           | does its polls even if its accuracy is higher. We wouldn't
           | necessarily expect it to keep outperforming other pollsters
           | in subsequent elections since the direction of polling bias
           | bounces around unpredictably from election to election.
           | 
           | - https://abcnews.go.com/538/best-pollsters-
           | america/story?id=1...
        
       | jdoliner wrote:
       | I don't know about the framing of "giving up." But I think anyone
       | who's been following election models since the original 538 in
       | 2008 has probably gotten the feeling that they have less alpha in
       | them than they did back then. I think there's some obvious
       | reasons for this that the forecasters would probably agree with.
       | 
       | The biggest one seems to be a case of Goodhart's Law, leading to
       | herding. Pollsters care a lot now about what their rating is in
       | forecasting models, so they're reluctant to publish outlier
       | results, those outlier results are very valuable for the models
       | but are likely to get a pollster punished in the ratings next
       | cycle.
       | 
       | Lots of changes to polling methods have been made due to polls
       | underestimating Trump. Polls have become like mini models unto
       | themselves. Due to their inability to poll a representative slice
       | of the population they try to correct by adjusting their results
       | to compensate for the difference between who they've polled and
       | the likely makeup of the electorate. This makes sense in theory,
       | but of course introduces a whole bunch of new variables that need
       | to be tuned correctly.
       | 
       | On top of all this is the fact that the process is very high
       | stakes and emotional with pollsters and modellers alike bringing
       | their own political biases and only being able to resist pressure
       | from political factions so much.
       | 
       | The analogy I kept coming back to watching election models during
       | this last cycle was that it looked like an ML model that didn't
       | have the data it needed to make good predictions and so was
       | making the safest prediction it could make given what it did
       | have. Basically getting stuck in this local minima at 50-50 that
       | was least likely to be off by a lot.
        
         | genewitch wrote:
         | okay, now change "election prediction models" to "Climate
         | models" and see if you feel like downvoting me merely for
         | pointing out the (slight?) hypocrisy in "excusing" every other
         | model we humans use for being "inaccurate" or "not having the
         | full details" or the "whole slice of"...
         | 
         | when none of the models tend to agree... and the IPCC
         | literature published however often they do it is hung upon the
         | framework of _models_.
        
           | sojournerc wrote:
           | All models are wrong, but some are useful.
           | 
           | Climate modeling is way messier than the media portrays, yet
           | even optimistic models show drastic change.
           | 
           | I'm not in the catastrophy camp, but it's worth preparing for
           | climate change regardless of origin. It's good for humanity
           | to be resilient to a hostile planet.
        
       | lyu07282 wrote:
       | At the end of the day it was a 312 vs. 226 in the Electoral
       | College. Seems a bit odd that this is supposed to be impossible
       | to be predicted with any useful amount of certaintly. But perhaps
       | that says more about the nature of the Electoral College than it
       | says about pollsters.
       | 
       | https://en.wikipedia.org/wiki/Efforts_to_reform_the_United_S...
        
         | culi wrote:
         | Almost every major model had this outcome as very likely.
         | 
         | There was an inordinate amount of very close swing states in
         | 2024. Whichever way the swing was gonna be was likely to apply
         | across the board.
         | 
         | If you're going by popular vote, then this was actually the 8th
         | closest election in all of US history.
        
       | alphazard wrote:
       | Highly recommend this video by NNT about why prediction markets
       | tend towards 50-50.
       | 
       | https://www.youtube.com/watch?v=YRvPF__du9w
       | 
       | Prediction markets are usually implemented as binary options.
       | Like vanilla options, their price depends not just on the most
       | likely outcome, but the whole distribution. When uncertainty
       | increases (imagine squishing a mound of clay), you end up pushing
       | lots of probability mass (clay) to the other side of the bet and
       | the expectation of the payoff (used to make a price) tends
       | towards 1/2.
        
       | moduspol wrote:
       | Agreed with the points in OP. Though we did have the story that
       | came out shortly after the election that apparently internal
       | polling for the Harris campaign never showed her ahead [1].
       | 
       | Obviously it says in the article that they did eventually fight
       | it to a dead heat, which is in-line with a 50-50 forecast, but I
       | do wonder what, if anything, failed such that this key detail was
       | never reported on publicly until after the election.
       | 
       | As the article notes, public polls started appearing in late
       | September showing Harris ahead, which they never saw internally.
       | Are internal polls just that much better than public ones? Is the
       | media just incentivized to report on the most interesting outcome
       | (a dead heat)?
       | 
       | [1]
       | https://www.usatoday.com/story/news/politics/elections/2024/...
        
       | delichon wrote:
       | There's a pretty clear test of whether a pollster is about
       | reporting or influencing: the partisanship of their errors.
       | Neutral pollsters have differences with the results randomly
       | distributed across parties. Propagandists have errors that favor
       | their patrons. This essay leans on the magnitude of the errors,
       | but that's less probative than their distribution.
       | 
       | Does any poll aggregator order by randomness of polling error?
        
         | culi wrote:
         | This mixes up accuracy with precision and 538 had written at
         | length about this.
         | 
         | There's a big difference between pollster bias correction vs
         | rating.
         | 
         | There are many pollsters that are pretty consistently +2D/R but
         | are reliably off from the final result in that direction. These
         | polls are actually extremely valuable once you make the
         | correction in your model. Meanwhile, polls that can be off from
         | the final result by a large amount but _average out_ to about
         | correct should not be trusted. This is a yellow flag for
         | herding
         | 
         | A pollster can have an A+ rating while having a very consistent
         | bias in one direction or another. The rating is meant to
         | capture consistency/honesty of methodology more than result
        
       | notfried wrote:
       | For each of the 2024 7 swing states, the winner was <1% ahead on
       | average, so what good are these polls if the results are going to
       | be within their margin of error?
       | 
       | They need to either find a more accurate way, or... give up!
        
         | culi wrote:
         | We had a pretty weird year in general. Harris did bad across
         | most safe states but seemed to do much better than her average
         | in swing states (not enough to win them, but much better than
         | she did in non-competitive states)
         | 
         | Many election models rely heavily on historical correlation.
         | States like OH and IN might vote quite differently but their
         | _swings_ tend to be in the same direction.
         | 
         | The weirdness this year (possibly caused by the Harris campaign
         | having a particularly strong ground game in swing states)
         | definitely challenged a lot of baked in assumptions of
         | forecasts.
        
         | rachofsunshine wrote:
         | What they're good for is telling you that things are close. A
         | tied poll or a 50-50 model can tell you that if your beliefs
         | think it's 99% to go one way, you're probably overconfident,
         | and should be more prepared for it to go the other way.
         | 
         | I cared about the result, because it was going to decide
         | whether I settled down in the US or whether I wanted to find a
         | different place to live. And because I paid attention to those
         | polls, I knew that what happened was not particularly unlikely.
         | I prepared early, and that enabled me to be ready to celebrate
         | my new year on a Europe-bound plane over the Atlantic.
         | 
         | A lot of people I know thought it couldn't happen. They ignored
         | the evidence in front of them, because it was distasteful to
         | them (just as it was to me). And they were caught flat-footed
         | in a way that I wasn't.
         | 
         | That's not the benefit of hindsight: I brought receipts. You
         | can see the 5,000 equally-likely outcomes I had at the start of
         | the night (and how they evolved as I added the vote coming in)
         | here:
         | https://docs.google.com/spreadsheets/d/11nn9y9fusd-6LQKCof3_...
         | .
        
       | randomNumber7 wrote:
       | Tell me you don't know the difference between bias and variance.
       | The longest article/comment wins.
        
       ___________________________________________________________________
       (page generated 2025-03-10 23:01 UTC)