As someone who is struggling to teach his youngest daughter about why the Mode (or any math for that matter....which hurts me deeply) matters, in the "real" (and I use that term in the loosest possible sense, whenever the current political environment is involved), thank you for this post. The modal outcome of the model. Very useful in life.
I just read this over and I realize I am giving a very unhelpful reply for someone looking for something "real" but I'mma post it anyway.
Not knowing what age your youngest daughter is but because the mode comes up seemingly as a new topic, I'm going to guess late elementary school. "The Cat in Numberland" is one of my go-to presents for parents of elementary-aged kids.
Infinity is WILD.
I guess my love for math came NOT from its usefulness, but from how purdy it can be.
I can't really distinguish between the most "hard-core" mathematicians I know and artists seeking out beauty with a very, very quirky medium.
If she likes space, math's for her. We know about space because of math.
If she likes music, math's for her. Math helps us understand harmonies and can help us discover new ones.
If she likes art or architecture, math's for her. Things stay up against the forces pushing to them because of calculations we make.
If she likes animals and cares about the environment, math's for her. We model populations of endangered and invasive species with math.
And of course, if she cares about politics, math's for her. But Nate already has that covered.
It's useful because it provides one more way to argue ex post facto that your model was right? I guess that's a use! Generally speaking, the mode is the least useful out of mean, median and mode.
I didn't mean for this to evolve into a math conversation. As a non-actuary, who has literally thousands of them as colleagues I have a lot of those! IMO each of the three have utility in certain situations. As non political example: Are teams still tanking in MLB? Well the mean record is useless (81 wins). The median is interesting. But, if the mode is say, and I'm making this up, 74 wins, and the median is also well under 81. Well... teams are tanking. If not, then the system is no longer rewarding tanking. Just one man (a man who loves parentheticals) opinion.
I am an actuary, but I'm not a baseball fan, so I have no idea what "tanking" is! If you are talking about skewness then comparing the mean and median will give you a rough indication of skewness without bringing the mode into it. The mode is particularly unhelpful in something like election modelling because you are dealing with a multimodal distribution. As Nate said in the article, the two most likely outcomes were the opposite extremes - Trump taking all the swing states or Harris taking all the swing states. The fact that Trump taking them all is slightly more likely than Harris taking them all really isn't a useful indication of anything.
But that's just because Harris had a significant chance of getting 6/7 of the swing states, whereas Trump was more all-or-nothing. If your main interest is in who actually wins the election, that isn't really very interesting.
Nate, your work is brilliant and I’m a big fan of your recent book. But, going forward, the value added in polling (and aggregators) to predict winners/losers, probabilistic intervals, or to determine relative percentage of vote shares needs to be reassessed. I have mad respect for your insights and skill. But, this type of model building is unworthy of your talents.
1. The large betting markets produced earlier and more accurate projections than polling or polling aggregators.
2. A miss of 2 percent or more in a swing state is not a good outcome. A swing state is by definition going to be close.
3. Polls underestimated Trump in 3 consecutive elections even though polling companies were doing all they could to correct this bias. They were unable to do this. The odds of the polling error in R favor in 3 consecutive elections would only occur in .5 x .5x .5=12.5 percent of the time. There appears to be no available fix.
4. The error in the national vote share will be 2.5-4.5 percent. Regretfully, the highest rated pollsters perform no better or even poorer than the less respected ones. Your model is not alone in this: the Economist and 538 will have the same level of error magnitude as S B.
5. A flip of the coin to determine the winner of every swing state (they will still likely be the same 7) in 2028 produces more accurate projections than polling or aggregators.
All the best my friend. I truly look forward to your next book.
> The error in the national vote share will be 2.5-4.5 percent. Regretfully, the highest rated pollsters perform no better or even poorer than the less respected ones.
Polymarket also had Harris as 70% likely to win the popular vote (Nate had her at 76%). They also had Harris at 60% to win Michigan the day before, on a pretty liquid market with $28M in volume: https://polymarket.com/event/michigan-presidential-election-winner (vs Nate at 61.8%). If Polymarket was truly superior, they'd have more accurate predictions for _every_ state, not just the overall prediction.
> The large betting markets produced earlier and more accurate projections than polling or polling aggregators.
On the night of the election? They did but... that's not the most important part of predicting elections? Beating "The Needle" by 30 minutes isn't all that impressive, the NYT is intentionally trying to be a little conservative on election night. I would also want to see Polymarket's performance in a harder election, such as 2020 when the situation was less clear and it took much longer to declare a winner.
I suspect the betting markets aren't very accurate on many things currently, and wouldn't have been in the aftermath of the 2020 election for example. But that seems to be because there are extraordinary restrictions on participation and so they aren't very responsive/efficient (eg they are markets open now on popular vote %-- because CA takes months to count its vote--that have around 30k invested) . If election betting becomes an open and efficient market, i would *bet* that the election modeling industry is gone. I suspect nate probably knows this, though i haven't read his book so i don't know how he feels about election betting markets.
Polymarket isn't available to U.S. bettors currently(nominally, at least), I think. And PredictIt and Kalshi have been hamstrung all year, watching a set of court cases with the SEC re betting markets. I personally find those cases pretty interesting, and I think the regulatory scheme for those markets is a pretty interesting question. But the point remains that if it ever becomes as easy to bet on the election as it is to buy Google stock, the wisdom of the masses will reign. (Not necessarily because people like Nate's talents will be useless, but because they will be infinitely MORE useful and valuable, and to people who are willing to pay much more than us Substack subcribers lol.)
I'm not sure that is entirely fair. The model correctly predicted that it would be a close race. You can't really expect it to do more than that in a close race. The model is more valuable when it isn't such a close race. For example, in 2016 the polling showed a pretty safe win for Clinton, but Nate's model correctly said that Trump was actually in with a very real chance - the model served a very useful purpose there. Those of us that listened to it were a lot less shocked by the result than those that just listened to the conventional wisdom.
Nate has always said that the purpose of the model is not to predict the result but to quantify the uncertainty in the result. It does that very well. It's not his fault that in this case the result was extremely uncertain.
BTW, don't quote your academic credentials on a post that is completely outside your field of expertise. Your post has nothing to do with business law or accountancy. It makes people take you less seriously, not more...
I mean, 70% chance of victory seems like "safe win" territory, no?
For example, this election should have been categorized a safe win for Trump- he swept all the blue wall states and won the popular vote. Nate having it be a 50/50 seems pretty off- it should have been 70/30 Trump or something similar to the betting markets.
I certainly wouldn't call 70% safe. That's a clear favourite, but not anywhere near safe. If you were playing a game and need to roll a 3 or higher on a dice to win, would you say that was a safe win? People roll 1's and 2's all the time.
The size of the win doesn't really tell you anything about the likelihood of the win. If there is a large margin of error, then you can get a strong win for either candidate despite the polling suggesting a dead heat.
>That's a clear favourite, but not anywhere near safe. If you were playing a game and need to roll a 3 or higher on a dice to win, would you say that was a safe win? People roll 1's and 2's all the time.
Yes I would absolutely take that as a safe bet.
>The size of the win doesn't really tell you anything about the likelihood of the win
In this situation it does - Trump winning the popular vote indicates that Harris hardly had a chance to begin with.
The issue with margin of error is that pollsters never made methodological changes to account for the underpolling Trump has had over the years.
You have a very different understanding of the word "safe" than is the ordinary usage.
Was the failure of pollsters to adequately address past problems known before the election, though? You can only allow for that in the probability if it was known about in advance.
> 5. A flip of the coin to determine the winner of every swing state (they will still likely be the same 7) in 2028 produces more accurate projections than polling or aggregators.
May you clarify this? A flip of the coin for all 7 swing states gives a probability of (1/2^7)\approx 0.7%, while Silver's model gave the probability of an R clean sweep 20%. If you're comparing it to a flip of the coin to determine an R clean sweep or a D clean sweep, that seems like an unfair comparison.
I haven't been convinced that pollsters were doing all they could to correct the error. And my intuition is that isn't very surprising. If WaPo has trump losing, what incentive is there really for NYT to publish a poll showing trump winning? (Losing subs to WaPa, likely.) Might as well miss together. I could be wrong, but no one has presented a satisfactory answer other than, "well, nate cohn says they're trying really hard."
The publications commissioning the polls don't have much incentive to be right. They want polls that produce interest stories, not polls that get the right answer. But they aren't doing the polling. They are just commissioning it. The pollsters have a strong incentive to get it right because they do political polling as a way to market their non-political polling. The more accurate they are, the more work they'll get from marketers and other consumers of private polling.
Most polling is by businesses for marketing purposes. They want to know if their marketing campaigns are working, for example, so they poll people to ask if they are aware of their business or product. Or they want to know which demographics they should be focusing their marketing on, so they poll people to find out who uses or is considering using their product.
I see, that's helpful. But, is private polling still a robust business model ? E.g., at a time when google knows most of my online browsing info and can sell that to businesses, along with some estimated demographic info? And is it really true that Siena College is making more from market research than they get from the NYT? That could be the case, because I don't know anything. These are just the thoughts of an uninformed skeptic.
Google doesn't collect information on your thoughts. Not yet, anyway! It doesn't know if you have actually noticed any of the ads it has shown you, for example. An advertising campaign that is completely forgettable doesn't work and polling is the best way to identify that.
Here is a list of examples of things Sienna College has done. It looks like they are focusing more on public sector than commercial, but that may be a bias in the projects clients let them share. The basic point holds, though - election polling is a very small part of what they do.
> "The large betting markets produced earlier and more accurate projections than polling or polling aggregators."
The large betting markets are heavily influenced by polling aggregator results.
You can't easily do a controlled experiment and see what the betting markets would have come up with when there are no polls or polling aggregators, but there is a significant likelihood the betting markets would then be way out.
Further, the betting markets are so small that it is easily possible for a determined billionaire to sway them.
> "Polls underestimated Trump in 3 consecutive elections even though polling companies were doing all they could to correct this bias. "
Sure, but the eventual outcome might just have reflected the fact that undecideds all broke for Trump because they just weren't ready for a woman president - or some such factor difficult to tease out.
Hopefully, asking the right questions in after the event polls might give a clue. Of course, there shouldn't be a 4th election involving Trump, so the pollsters may not believe it is worth their while to identify specific factors.
There is a pretty good chance that the 2016 error resulted from Comey, FBI head, putting his finger on the scale by announcing he had another email server with Clinton emails on it, and suppressing the fact he had an intelligence briefing document on Trump.
But on the reverse, it is difficult to see any mechanism whereby the polls and polling aggregation would be influenced by the betting markets. You could argue that rich whales betting on Trump also donated considerably to the Trump campaign, but there are published figures showing donations to Harris were higher than to Trump.
> "A miss of 2 percent or more in a swing state is not a good outcome. A swing state is by definition going to be close."
2% in the polls for a state seems pretty good - well below the statistical uncertainly in individual polls of a few hundred voters in a state. And we know for sure there was a lot of herding going on, likely creating systematic errors, but previously didn't know in what direction.
> "The error in the national vote share will be 2.5-4.5 percent. "
"Democrats, however — and here, I’m not referring so much Silver Bulletin subscribers but in the broader universe online — often get angry with you when you only halfway agree with them. And I really think this difference in personality profiles tells you a little something about why Trump won: Trump was happy to take on all comers, whereas with Democrats, disagreement on any hot-button topic (say, COVID school closures or Biden’s age) will have you cast out as a heretic. That’s not a good way to build a majority, and now Democrats no longer have one."
I am a Reagan republican, and old enough to have been cast from the original mold. Your comments are good to see from a Harris supporter. I am convinced, due to Covid and the over reaction to Floyd, that "normal America" - those who do not believe in groupthink, men playing girls' sports, and don't want to be "burdened" with a thesaurus of pronouns or the responsibility of medically mutilating children - embraced an opportunity to snap the pendulum back the other direction. It is hardly surprising that they found common cause with Latinos who tend to be catholic, family oriented, and committed to hard work. They too, at least in my experience in Texas, tend not to obsess over unknown opinions on social media.
Assuming republicans have the sense to maintain the momentum of the multi-ethnic working class coalition, this is an enduring movement that can continue well beyond a Trump presidency. So long as democrats continue to focus on things like messaging, race, gender, or candidates rather than policy and social class, republicans will have little competition in effectively engaging those voters.
As a Brit, this reminds me of our election in 2019 which Boris johnson (our trump) won in a fairly big landslide. Everyone thought that was it for ever for the Labour Party that had become weighed down by idealistic agitators obsessed with identity politics. But five years later it’s the Conservative Party (our republicans) facing the existential crisis. Johnson turned out to be mildly corrupt and very incompetent and his successors much worse. And the Labour Party course corrected to make itself electable (they aren’t exactly exciting but they focused on just being boring and competent). They will still be in power after trump leaves office assuming he doesn’t make himself king.
Critically imho the main reason trump won is inflation. All incumbent governments have been hammered in the polls this year because people are feeling much poorer than they did three years ago. But trumps policies are inflationary: so either he drops them or the republicans will be toast.
Thanks for sharing this opinion- I’m definitely curious about the efforts the Labor party made to be more electable. Are you comfortable sharing what you think those efforts were?
In the disability community I’m part of, there are huge concerns about the UK’s push towards “right to work” and implied cuts to disability benefits (or higher scrutiny) for those who simply cannot work. However I can also see how such a policy would be popular among working class voters. Does that match your read?
It will be interesting to see if the Dems can course correct. The activists are the really extreme ones, not so much the Dem base. But they have outsized influence.
This is a well argued interpretation of the electorate’s sentiments, but I see no evidence that this rightward shift is “an enduring movement that can continue well beyond a Trump presidency.” The movement is based on the cult of personality of one person, and a majority of voters were willing to look past Trump denying the results of the last election and encouraging his supporters to storm the Capitol and hang his own Vice President. You’re assuming that Trump will act rationally and try to build a legacy that will endure well beyond his second term. Instead, we will get more of what we’ve always seen from him — petty grievances, unpredictability, stoking divisions, alienating close allies, and rage. How can that legacy possibly endure? Seventy million Americans can’t stand him and never will.
And you miss my point completely. Yes, there is core Trump constituency. But the reasons I noted above have nothing to do with Trump per se, but represent the basis for an enduring and successful opposition to the insanity which the democrat party has come to represent. That has nothing to do with Trump's faults but everything to do with an ever more radical ideology that is the antithesis of "normal" America. Trump did better everywhere this year in spite of the faults you listed - not because of them.
The democrat party discounts that to its lasting political peril. JD Vance or Ron DeSantis, coupled with the right running mate, will be a far more compelling figures than Trump in four years further solidifying a multi-ethnic working and middle class base that has the potential to only grow.
One other little point. Another huge motivating factor for many who supported Trump in spite of his flaws (I am one of them), is the solidly held belief that the democrat party coupled with traditional media represents a far greater threat to democracy that Trump ever has.
I agree with you on a lot of your diagnoses of why the democrats lost and why they will continue to lose if they do not address. However, you have asserted that the democrat party represents a far greater threat to democracy than Trump ever has without any evidence. And we know that that which can be asserted without evidence can be dismissed without evidence, but I will nonetheless provide you with evidence that clearly and obviously shuts down your argument. The democratic process involves the people voting and those results being enacted. The people of Georgia, Arizona, and Pennsylvania decisively voted to elect Joe Biden, yet Trump made efforts to disenfranchise those citizens--an action that would have effectively killed democracy in the 2020 election--to have their votes thrown out and unilaterally changed to the exact opposite option vs what they selected. Not to mention Trump's involvement in fomenting J6. What would J6 have looked like if it was successful? Would Pence have been held hostage until he selected the false delegates and unilaterally disenfranchised American citizens by installing a president in direct contrast to the candidate selected through the prescribed democratic process that we all engage in selecting representatives under? You demonstrate extreme cognitive dissonance by ignoring these indisputable facts and assert, again, without any evidence, that democrats are a bigger threat to democracy than Trump ever has been. You claim this without evidence because there is no credible evidence available for you to present.
I agree it's likely (fervently hope so, actually) that what you call "normal America" will create an enduring movement that embraces free speech over enforced left-wing correctness. However, I don't know why that movement needs to include the more toxic parts of MAGA culture like the relentless misogyny, opposition to reproductive rights, enforced Christianity, desire to ban all LGBT not just youth gender transition, love of foreign dictators and authoritarianism, making health insurance harder to get again, etc.
I have been thinking, only half in jest, that I should switch from being a Democrat to being a never-Trump Republican (moderate Eisenhower/Weld/Baker side, to be sure). If either party manages to resurrect the moderate center, that would be a movement that could endure.
I actually think that is an absolutely reasonable position to take Monica. I would quibble about things like relentless misogyny (take a week and pull up Megyn Kelly's podcasts of her radio program - I mean that seriously). You will be hard pressed to find a stronger, more able, or intelligent woman on either side of the aisle. But she also prefers her men to be men while she is being a woman. Also, it is probably worth noting that Trump did just choose a very tough and competent woman to serve as his chief of staff - something no other president has done. And lest we forget, it was he who nominated Nikki Haley to be UN Ambassador during his first term. For whatever it is worth, she was my choice for the nomination.
Reproductive rights are a difficult issue. I do believe much majority opinion is regionally based. In other words, it is entirely possible, even reasonable, that respective majorities of women would have different opinions on the issue in different parts of the country. Hence, I think the "leave it to the states to sort it out" solution is extremely wise. We conveniently forget that Justice Ruth Bader Ginsburg almost certainly would have agreed.
But it is a tricky subject. I do not think any serious person disagrees with the issues of rape, incest, or the health of the mother. But there is a reasonable question that needs to be answered. At what point does the state have a responsibility for the health and safety of the unborn child? I think a very large majority of Americans, male or female, would find an issue with unnecessarily aborting a fetus in the third trimester, much less killing an aborted child that somehow survived the procedure. I also think most pregnant women would expect the state to aggressively charge and prosecute any second party who deliberately or negligently harmed an unborn child in the womb. So how do we resolve those conflicting demands? I do not pretend to know the answer to that, but I trust the wisdom of the people - particularly where they reside - far more than the Washington bureaucracy.
Everyone's been dunking on Silver Bulletin and 538 for being less accurate than prediction markets, but the per-state data shows that traditional modeling was still superior.
that's a strange criticism, since polling aggregators are of course not exogenous to prediction markets whereas the reverse can't be true without leading to feedback loops, as Nate noted in his Election Day SBSQ:
> Do you anticipate ever assigning “pollster ratings” to prediction markets and incorporating them in the model?
> No. Because I worry about recursive effects. Say our model accounts for prediction markets, so it moves toward — in this election — Trump. Then prediction markets see that and say “wow, Nate’s model is moving toward Trump!”. So Trump’s price goes up further. Then the model sees that, and Trump’s probability improves even more. And so forth.
That's a strange criticism of my criticism. Yes, prediction markets are subject to feedback from Nate's model as well as all kinds of other problems. But my point is that their accuracy isn't superior to traditional analytical methods just yet, as evidenced by looking at the sub-markets instead of just the headline market. *Why* that's the case doesn't matter, we're just discussing what numbers both models provided.
And this is why Polymarket's had to use a trick to claim they've predicted all 50 states correctly, instead of just taking the numbers from Nov 4th.
With respect, the betting models have a better track record for calling winner (90 percent or more). The models have missed the last 2 of 3. It’s not the modelers fault. Pollsters are not providing the models with worthy data.
"Admittedly, the sorts of Republicans who encounter our work are not a representative sample, probably being on the moderate side"
I would wager this is not correct. I am a subscriber and I am to the right of Genghis Khan. The "moderate" Republicans I know are mostly checked out and do not pay attention to news at all. They are certainly not committed enough to look at polls and models every day.
I wouldn't even call myself a Republican, but I did vote Trump, and I'm fairly moderate in that I support national health care, am concerned about climate change, support mandatory vaccines, want FOSTA/SESTA repealed, want a child tax credit, etc.
Immigration, wokeness/identity politics, cancel culture, gun rights, free speech, corporate free speech, the smugness and condescension of the left, the scandal regarding hiding Biden's senility, religious freedom...
Amazing how Trump gets free ride on felonies and sexual assault and general vulgarity. And Harris, the product of public schools and broken family is dinged on free speech, smugness, condescension, and... Scandal LOL! Scandal of all things to ding Harris. Wow America. Really? You're Against scandal so vote Trump? Let other generations worry about climate change. Let the kids worry about social security. Let Trump and the brilliant RFK take away pre existing conditions protection.
"I don't believe he ever committed sexual assault. The accusations are years or decades old and not particularly credible." Forgive me if I discredit your opinion on this given that the jury of 12 of Trump's peers unanimously agreed that they believe he did when presented evidence by both Trump and the plaintiff and your lack of knowledge on the matter relatively. When one considers that your level of analysis of this matter is about on par with that of someone with median intelligence, it is extremely discouraging to consider that half of the population will have an even poorer assessment of the situation.
Hahaha he's literally on audio admitting sexual assault. Degeneracy is so refreshing! Court decisions are illegitimate! Criminality isn't a bug, it's a feature! Clearly we've got the government we deserve.
There is no "global warming." That is all nonsense designed to switch the public out of oil and coal energy and into renewables. Which I would be fine with if it's organic and not subsidized and forced, but I sure don't like the Big Lie of global warming that seems to go on forever, with never any actual evidence.
The Supreme Court ruled that people could join together to form organizations to express their views, in particular political speech, without restrictions on funding. This was a great First Amendment case called Citizens United.
The Dems are against it and have campaigned on overruling it, which is truly scary, because that would mean banning political books or movies before an election.
We have a system similar to what the Dems are advocating for in Canada (severe limits on third party spending to prevent PAC-style advertising combined with caps on donations to force campaigns to use only small dollar donations) and the result has been to essentially silence large chunks of the political spectrum in the run-up to elections. Ironically, the consensus is that it has actually harmed left-wing political parties more, because they're more dependent on fundraising/spending from civil society organizations, activist groups, and unions, as opposed to right-wing parties which do better raising moderate sums from individuals.
I would honestly argue the opposite. I’m extremely against citizens united and think that the unlimited flow of money into politics has let the rich basically buy our politicians. They don’t work for us anymore (or at least even less than they used to) they work for the highest bidder.
I’m not sure why you say this would lead to political books and movies being banned. We had campaign finance limits for over a hundred years up until 2008 and this didn’t happen.
I think Nate is hedging his bets on his ability to attract conservative Trump supporters, but he's underestimating his appeal. There's many Trump supporters here, often to the confusion and anger of some liberal subscribers.
He's probably assuming conservatives wouldn't want to subscribe to him because he's a liberal, but he's forgetting that we only care about merit and not identity. That's kind of our thing. He's very good at his job so here we are. I still feel that way even if the model was a little off. It's dependent on the polls that are fed into it after all. If anyone dropped the ball it was the pollsters, not Nate. Garbage in and garbage out.
Hi Nate, can you do us a huge favor, before winding down the election content...
FRENCH WHALE STORY PLEASE.
Is the 'neighbor' question not herded in the original polls, just because the neighbor question isn't in mainstream news? What's the veracity of this stuff? Has this been reported all along? Will this ever end up in the model, or is it already?
Apparently the whale did his own polling that asked who people thought their neighbors would vote for, and the result was overwhelmingly for Trump. 'French whale' is a total misnomer for anyone doing their own poll, and especially for someone who makes a ton of money and then walks away from the table.
Polling error inside the margin is nothing to be ashamed about. I'm impressed with the correlated outcomes by state, which are less attributable to model input. I'm most interested in more analysis on the LV national polling miss and what the polling industry is planning to do to align sample size by region and close the LV/RV gap.
Save for most likely engaging in controlled herding in the future
and *maybe* offering more cross tabs to end clients earlier.
Which seems counter productive, but this is where I always have to tell people that 'Political Polling is The World Poker Tournament for Polls, you mostly come to be seen, not to win'.
No one makes real money (and often don't even break even) doing these political polls, but you can find good regional offices and get your name up where corporate clients will see, so you do it anyway.
The poll margin, *in the regions that most pollsters specialize in*, were very good. Outside those areas, there was variance. So, in the end, they aren't going to benefit from investing.
Add in that the next election will be a completely different ball game, with no Clinton Era holdovers and Trump closing out his Lame Duck term, and you could over-correct your company into oblivion. So, best for most of them to play the safe game.
If a hundred people take a random guess some of them are going to be right but it's not necessarily more useful than having a model that quantifies the uncertainty and enumerates the possible outcomes and their relative likelihood.
Trump did overperform a lot on the national popular vote but that was mostly due to overperformance in deep blue states where there wasn't a lot of polling (due to them not being seen as being in play). It's not surprising that the model didn't catch that, given that its prediction of the national popular vote is mostly extrapolated from state-level polling.
I wouldn't go that far - I think this certainly points to a weakness of extrapolating the state-level polling into national-level results. Maybe the imputed variance due to lack of polling should increase or decrease until the national popular vote distribution from the simulations matches that derived from the national-level polling?
And yes I understand that means she was expected to lose the popular vote 20% of the time which is not all that rare. But I suspect this was the model not working (or bad polls) and not just a fairly unlikely thing happening.
No, a "grift" would be Nate selling a thing, while claiming the thing is something else.
This is a statistical model designed to predict the outcome of elections based on polling, and Nate has sold it as just that. It's pretty obvious to me that he has put a lot of effort and his expertise into it.
What you've paid Nate for is access to the outputs of this model, plus his commentary on it (up to and including his own punditry).
What on earth else were you expecting from your subscription?
I actually think this specifically *was* an issue with Nate’s model rather than polling. Plenty of high quality national polls showed Trump tied or ahead of Harris in the final weeks. I think Nate has said that his model didn’t use national polling much, but instead extrapolated the national outcome from state polling. It seems like that would be self-reinforcing of the idea that there would be a big gap between the popular vote and electoral college outcome in favor of republicans, as was true in the last two elections, but it looks like that wasn’t really the case in the actual election outcome.
I didn't get the sense you were blaming Nate, and yeah you're absolutely right.
Honestly, I think people expect a bit too much from polls. When sampling error alone (setting aside methodological error) is 3-4pts *for each candidate*, and people expect you to predict the outcome in a state when both candidates are hovering in the 45-50 range... people are not gonna like the result.
And then I also guess with partisanship every election is going to be close. Polling misses seem to be systematic and not state by state so it just seems like a tremendous level of analysis to say polls are close but they could miss j one direction and equally likely in another. If the polls and state outcomes were less correlated perhaps the model could provide more insight and do more useful probability crunching.
They're absolutely falsifiable over the long term, just look at multiple predictions.
If you systematically looked at 100 80% predictions and the model only got them right 60% of the time, the model probably isn't very good. If it gets them right 75% to 85% of the time, it looks much better. If it gets 100% of those predictions right, the model is actually bad because it's way too under-confident.
National presidential elections are rare, but the model makes way more predictions than just the outcome of the presidential election.
Every model he runs is different. So you can’t compare some basketball model to the presidential election one. It would take literally millennium to have a data set where you can say okay there were 10 80% elections and the favorite one 8 out of 10 times. It is 100% something that cannot really be proven wrong.
You don't need 10 80% elections and 10 70% elections on so on to evaluate the model's calibration.
Instead, imagine that you can bet against the model. On each probabilistic forecast, you bet $1, and the model pays out $1/p where p is the final probability of the event you're betting on (so for an 80% model output, it pays $1.25). You add up the results of these bets, $1 on each forecast, and in the long run the model is well-calibrated if it returns $0.
You can use the other political models, though - congress, senate, governors. They all use the same basic approach as the presidential one. It's not a perfect assessment of the model, but it's pretty good. You can't really assess the model when it is saying the race is 50/50, though - Nate made that point himself a week or two ago. A model that always says it is 50/50 will come out as perfectly calibrated.
To the extent that's true, why has the actual outcome vs. his prediction always been well towards the tail, IN THE SAME DIRECTION, each and every time Trump has been on the ballot? If the model were unbiased, getting, say, a 20% tail result in the same direction three times in a row would happen ~0.8% of the time.
So now we're down to priors: Nate's model is great as is and we just need to have Trump on the ballot a bunch more times and it will all even out, OR...
"To the extent that's true, why has the actual outcome vs. his prediction always been well towards the tail, IN THE SAME DIRECTION, each and every time Trump has been on the ballot?"
I think there are several reasons for this, but the main reason is that the model is based primarily on polling, and polling has underestimated Trump consistently. You can build a model that operates on the assumption that polls will consistently underestimate a candidate, but then it's not really a polling-based model anymore.
You can make the argument that some polls should've been weighted more heavily than others, based on the results of this recent Presidential election. And that argument may now be (given the results) evidence-based, and will probably be incorporated into the next model (if there is one). But that argument wasn't based on much evidence previously.
"20% tail result in the same direction three times in a row would happen ~0.8% of the time."
This only holds if the predictions are statistically independent. I don't think that can be assumed here. Way too much going on.
As someone who is struggling to teach his youngest daughter about why the Mode (or any math for that matter....which hurts me deeply) matters, in the "real" (and I use that term in the loosest possible sense, whenever the current political environment is involved), thank you for this post. The modal outcome of the model. Very useful in life.
You sir are a cut above the rest,
If only you could teach the rest of the 'adult' world the same lesson...
I just read this over and I realize I am giving a very unhelpful reply for someone looking for something "real" but I'mma post it anyway.
Not knowing what age your youngest daughter is but because the mode comes up seemingly as a new topic, I'm going to guess late elementary school. "The Cat in Numberland" is one of my go-to presents for parents of elementary-aged kids.
Infinity is WILD.
I guess my love for math came NOT from its usefulness, but from how purdy it can be.
I can't really distinguish between the most "hard-core" mathematicians I know and artists seeking out beauty with a very, very quirky medium.
If she likes space, math's for her. We know about space because of math.
If she likes music, math's for her. Math helps us understand harmonies and can help us discover new ones.
If she likes art or architecture, math's for her. Things stay up against the forces pushing to them because of calculations we make.
If she likes animals and cares about the environment, math's for her. We model populations of endangered and invasive species with math.
And of course, if she cares about politics, math's for her. But Nate already has that covered.
It's useful because it provides one more way to argue ex post facto that your model was right? I guess that's a use! Generally speaking, the mode is the least useful out of mean, median and mode.
I didn't mean for this to evolve into a math conversation. As a non-actuary, who has literally thousands of them as colleagues I have a lot of those! IMO each of the three have utility in certain situations. As non political example: Are teams still tanking in MLB? Well the mean record is useless (81 wins). The median is interesting. But, if the mode is say, and I'm making this up, 74 wins, and the median is also well under 81. Well... teams are tanking. If not, then the system is no longer rewarding tanking. Just one man (a man who loves parentheticals) opinion.
I am an actuary, but I'm not a baseball fan, so I have no idea what "tanking" is! If you are talking about skewness then comparing the mean and median will give you a rough indication of skewness without bringing the mode into it. The mode is particularly unhelpful in something like election modelling because you are dealing with a multimodal distribution. As Nate said in the article, the two most likely outcomes were the opposite extremes - Trump taking all the swing states or Harris taking all the swing states. The fact that Trump taking them all is slightly more likely than Harris taking them all really isn't a useful indication of anything.
20% vs 14%. It does seem interesting considering the race was 50/50.
But that's just because Harris had a significant chance of getting 6/7 of the swing states, whereas Trump was more all-or-nothing. If your main interest is in who actually wins the election, that isn't really very interesting.
According to Nate’s data Trump had a 10.4% chance to get 6 out of 7 and Kamala had a 5.8% chance.
what in the parenthetical hell are you trying to say old man
Nate, your work is brilliant and I’m a big fan of your recent book. But, going forward, the value added in polling (and aggregators) to predict winners/losers, probabilistic intervals, or to determine relative percentage of vote shares needs to be reassessed. I have mad respect for your insights and skill. But, this type of model building is unworthy of your talents.
1. The large betting markets produced earlier and more accurate projections than polling or polling aggregators.
2. A miss of 2 percent or more in a swing state is not a good outcome. A swing state is by definition going to be close.
3. Polls underestimated Trump in 3 consecutive elections even though polling companies were doing all they could to correct this bias. They were unable to do this. The odds of the polling error in R favor in 3 consecutive elections would only occur in .5 x .5x .5=12.5 percent of the time. There appears to be no available fix.
4. The error in the national vote share will be 2.5-4.5 percent. Regretfully, the highest rated pollsters perform no better or even poorer than the less respected ones. Your model is not alone in this: the Economist and 538 will have the same level of error magnitude as S B.
5. A flip of the coin to determine the winner of every swing state (they will still likely be the same 7) in 2028 produces more accurate projections than polling or aggregators.
All the best my friend. I truly look forward to your next book.
Tom Moore PhD, JD
Professor Emeritus
Georgia College
> The error in the national vote share will be 2.5-4.5 percent. Regretfully, the highest rated pollsters perform no better or even poorer than the less respected ones.
Polymarket also had Harris as 70% likely to win the popular vote (Nate had her at 76%). They also had Harris at 60% to win Michigan the day before, on a pretty liquid market with $28M in volume: https://polymarket.com/event/michigan-presidential-election-winner (vs Nate at 61.8%). If Polymarket was truly superior, they'd have more accurate predictions for _every_ state, not just the overall prediction.
> The large betting markets produced earlier and more accurate projections than polling or polling aggregators.
On the night of the election? They did but... that's not the most important part of predicting elections? Beating "The Needle" by 30 minutes isn't all that impressive, the NYT is intentionally trying to be a little conservative on election night. I would also want to see Polymarket's performance in a harder election, such as 2020 when the situation was less clear and it took much longer to declare a winner.
I suspect the betting markets aren't very accurate on many things currently, and wouldn't have been in the aftermath of the 2020 election for example. But that seems to be because there are extraordinary restrictions on participation and so they aren't very responsive/efficient (eg they are markets open now on popular vote %-- because CA takes months to count its vote--that have around 30k invested) . If election betting becomes an open and efficient market, i would *bet* that the election modeling industry is gone. I suspect nate probably knows this, though i haven't read his book so i don't know how he feels about election betting markets.
It might happen in the future but it didn't happen yet, despite Polymarket's hype.
Polymarket isn't available to U.S. bettors currently(nominally, at least), I think. And PredictIt and Kalshi have been hamstrung all year, watching a set of court cases with the SEC re betting markets. I personally find those cases pretty interesting, and I think the regulatory scheme for those markets is a pretty interesting question. But the point remains that if it ever becomes as easy to bet on the election as it is to buy Google stock, the wisdom of the masses will reign. (Not necessarily because people like Nate's talents will be useless, but because they will be infinitely MORE useful and valuable, and to people who are willing to pay much more than us Substack subcribers lol.)
I'm not sure that is entirely fair. The model correctly predicted that it would be a close race. You can't really expect it to do more than that in a close race. The model is more valuable when it isn't such a close race. For example, in 2016 the polling showed a pretty safe win for Clinton, but Nate's model correctly said that Trump was actually in with a very real chance - the model served a very useful purpose there. Those of us that listened to it were a lot less shocked by the result than those that just listened to the conventional wisdom.
Nate has always said that the purpose of the model is not to predict the result but to quantify the uncertainty in the result. It does that very well. It's not his fault that in this case the result was extremely uncertain.
BTW, don't quote your academic credentials on a post that is completely outside your field of expertise. Your post has nothing to do with business law or accountancy. It makes people take you less seriously, not more...
I mean, 70% chance of victory seems like "safe win" territory, no?
For example, this election should have been categorized a safe win for Trump- he swept all the blue wall states and won the popular vote. Nate having it be a 50/50 seems pretty off- it should have been 70/30 Trump or something similar to the betting markets.
I certainly wouldn't call 70% safe. That's a clear favourite, but not anywhere near safe. If you were playing a game and need to roll a 3 or higher on a dice to win, would you say that was a safe win? People roll 1's and 2's all the time.
The size of the win doesn't really tell you anything about the likelihood of the win. If there is a large margin of error, then you can get a strong win for either candidate despite the polling suggesting a dead heat.
>That's a clear favourite, but not anywhere near safe. If you were playing a game and need to roll a 3 or higher on a dice to win, would you say that was a safe win? People roll 1's and 2's all the time.
Yes I would absolutely take that as a safe bet.
>The size of the win doesn't really tell you anything about the likelihood of the win
In this situation it does - Trump winning the popular vote indicates that Harris hardly had a chance to begin with.
The issue with margin of error is that pollsters never made methodological changes to account for the underpolling Trump has had over the years.
You have a very different understanding of the word "safe" than is the ordinary usage.
Was the failure of pollsters to adequately address past problems known before the election, though? You can only allow for that in the probability if it was known about in advance.
A 70% win rate is superb- it's not 95% but most bets aren't.
Yes we've known for years about pollsters underrepresenting Trump voters and I haven't seen anything to indicate that they changed that.
> 5. A flip of the coin to determine the winner of every swing state (they will still likely be the same 7) in 2028 produces more accurate projections than polling or aggregators.
May you clarify this? A flip of the coin for all 7 swing states gives a probability of (1/2^7)\approx 0.7%, while Silver's model gave the probability of an R clean sweep 20%. If you're comparing it to a flip of the coin to determine an R clean sweep or a D clean sweep, that seems like an unfair comparison.
I haven't been convinced that pollsters were doing all they could to correct the error. And my intuition is that isn't very surprising. If WaPo has trump losing, what incentive is there really for NYT to publish a poll showing trump winning? (Losing subs to WaPa, likely.) Might as well miss together. I could be wrong, but no one has presented a satisfactory answer other than, "well, nate cohn says they're trying really hard."
The publications commissioning the polls don't have much incentive to be right. They want polls that produce interest stories, not polls that get the right answer. But they aren't doing the polling. They are just commissioning it. The pollsters have a strong incentive to get it right because they do political polling as a way to market their non-political polling. The more accurate they are, the more work they'll get from marketers and other consumers of private polling.
That's interesting. What do you mean by private / non-political polling? (I don't know much about the industry.)
Most polling is by businesses for marketing purposes. They want to know if their marketing campaigns are working, for example, so they poll people to ask if they are aware of their business or product. Or they want to know which demographics they should be focusing their marketing on, so they poll people to find out who uses or is considering using their product.
I see, that's helpful. But, is private polling still a robust business model ? E.g., at a time when google knows most of my online browsing info and can sell that to businesses, along with some estimated demographic info? And is it really true that Siena College is making more from market research than they get from the NYT? That could be the case, because I don't know anything. These are just the thoughts of an uninformed skeptic.
Google doesn't collect information on your thoughts. Not yet, anyway! It doesn't know if you have actually noticed any of the ads it has shown you, for example. An advertising campaign that is completely forgettable doesn't work and polling is the best way to identify that.
Here is a list of examples of things Sienna College has done. It looks like they are focusing more on public sector than commercial, but that may be a bias in the projects clients let them share. The basic point holds, though - election polling is a very small part of what they do.
https://scri.siena.edu/sample-projects/
> "The large betting markets produced earlier and more accurate projections than polling or polling aggregators."
The large betting markets are heavily influenced by polling aggregator results.
You can't easily do a controlled experiment and see what the betting markets would have come up with when there are no polls or polling aggregators, but there is a significant likelihood the betting markets would then be way out.
Further, the betting markets are so small that it is easily possible for a determined billionaire to sway them.
> "Polls underestimated Trump in 3 consecutive elections even though polling companies were doing all they could to correct this bias. "
Sure, but the eventual outcome might just have reflected the fact that undecideds all broke for Trump because they just weren't ready for a woman president - or some such factor difficult to tease out.
Hopefully, asking the right questions in after the event polls might give a clue. Of course, there shouldn't be a 4th election involving Trump, so the pollsters may not believe it is worth their while to identify specific factors.
There is a pretty good chance that the 2016 error resulted from Comey, FBI head, putting his finger on the scale by announcing he had another email server with Clinton emails on it, and suppressing the fact he had an intelligence briefing document on Trump.
But on the reverse, it is difficult to see any mechanism whereby the polls and polling aggregation would be influenced by the betting markets. You could argue that rich whales betting on Trump also donated considerably to the Trump campaign, but there are published figures showing donations to Harris were higher than to Trump.
> "A miss of 2 percent or more in a swing state is not a good outcome. A swing state is by definition going to be close."
2% in the polls for a state seems pretty good - well below the statistical uncertainly in individual polls of a few hundred voters in a state. And we know for sure there was a lot of herding going on, likely creating systematic errors, but previously didn't know in what direction.
> "The error in the national vote share will be 2.5-4.5 percent. "
Again, this is not unreasonable.
Well said, professor!
"Democrats, however — and here, I’m not referring so much Silver Bulletin subscribers but in the broader universe online — often get angry with you when you only halfway agree with them. And I really think this difference in personality profiles tells you a little something about why Trump won: Trump was happy to take on all comers, whereas with Democrats, disagreement on any hot-button topic (say, COVID school closures or Biden’s age) will have you cast out as a heretic. That’s not a good way to build a majority, and now Democrats no longer have one."
I am a Reagan republican, and old enough to have been cast from the original mold. Your comments are good to see from a Harris supporter. I am convinced, due to Covid and the over reaction to Floyd, that "normal America" - those who do not believe in groupthink, men playing girls' sports, and don't want to be "burdened" with a thesaurus of pronouns or the responsibility of medically mutilating children - embraced an opportunity to snap the pendulum back the other direction. It is hardly surprising that they found common cause with Latinos who tend to be catholic, family oriented, and committed to hard work. They too, at least in my experience in Texas, tend not to obsess over unknown opinions on social media.
Assuming republicans have the sense to maintain the momentum of the multi-ethnic working class coalition, this is an enduring movement that can continue well beyond a Trump presidency. So long as democrats continue to focus on things like messaging, race, gender, or candidates rather than policy and social class, republicans will have little competition in effectively engaging those voters.
As a Brit, this reminds me of our election in 2019 which Boris johnson (our trump) won in a fairly big landslide. Everyone thought that was it for ever for the Labour Party that had become weighed down by idealistic agitators obsessed with identity politics. But five years later it’s the Conservative Party (our republicans) facing the existential crisis. Johnson turned out to be mildly corrupt and very incompetent and his successors much worse. And the Labour Party course corrected to make itself electable (they aren’t exactly exciting but they focused on just being boring and competent). They will still be in power after trump leaves office assuming he doesn’t make himself king.
Critically imho the main reason trump won is inflation. All incumbent governments have been hammered in the polls this year because people are feeling much poorer than they did three years ago. But trumps policies are inflationary: so either he drops them or the republicans will be toast.
Thanks for sharing this opinion- I’m definitely curious about the efforts the Labor party made to be more electable. Are you comfortable sharing what you think those efforts were?
In the disability community I’m part of, there are huge concerns about the UK’s push towards “right to work” and implied cuts to disability benefits (or higher scrutiny) for those who simply cannot work. However I can also see how such a policy would be popular among working class voters. Does that match your read?
It will be interesting to see if the Dems can course correct. The activists are the really extreme ones, not so much the Dem base. But they have outsized influence.
This is a well argued interpretation of the electorate’s sentiments, but I see no evidence that this rightward shift is “an enduring movement that can continue well beyond a Trump presidency.” The movement is based on the cult of personality of one person, and a majority of voters were willing to look past Trump denying the results of the last election and encouraging his supporters to storm the Capitol and hang his own Vice President. You’re assuming that Trump will act rationally and try to build a legacy that will endure well beyond his second term. Instead, we will get more of what we’ve always seen from him — petty grievances, unpredictability, stoking divisions, alienating close allies, and rage. How can that legacy possibly endure? Seventy million Americans can’t stand him and never will.
And you miss my point completely. Yes, there is core Trump constituency. But the reasons I noted above have nothing to do with Trump per se, but represent the basis for an enduring and successful opposition to the insanity which the democrat party has come to represent. That has nothing to do with Trump's faults but everything to do with an ever more radical ideology that is the antithesis of "normal" America. Trump did better everywhere this year in spite of the faults you listed - not because of them.
The democrat party discounts that to its lasting political peril. JD Vance or Ron DeSantis, coupled with the right running mate, will be a far more compelling figures than Trump in four years further solidifying a multi-ethnic working and middle class base that has the potential to only grow.
One other little point. Another huge motivating factor for many who supported Trump in spite of his flaws (I am one of them), is the solidly held belief that the democrat party coupled with traditional media represents a far greater threat to democracy that Trump ever has.
I agree with you on a lot of your diagnoses of why the democrats lost and why they will continue to lose if they do not address. However, you have asserted that the democrat party represents a far greater threat to democracy than Trump ever has without any evidence. And we know that that which can be asserted without evidence can be dismissed without evidence, but I will nonetheless provide you with evidence that clearly and obviously shuts down your argument. The democratic process involves the people voting and those results being enacted. The people of Georgia, Arizona, and Pennsylvania decisively voted to elect Joe Biden, yet Trump made efforts to disenfranchise those citizens--an action that would have effectively killed democracy in the 2020 election--to have their votes thrown out and unilaterally changed to the exact opposite option vs what they selected. Not to mention Trump's involvement in fomenting J6. What would J6 have looked like if it was successful? Would Pence have been held hostage until he selected the false delegates and unilaterally disenfranchised American citizens by installing a president in direct contrast to the candidate selected through the prescribed democratic process that we all engage in selecting representatives under? You demonstrate extreme cognitive dissonance by ignoring these indisputable facts and assert, again, without any evidence, that democrats are a bigger threat to democracy than Trump ever has been. You claim this without evidence because there is no credible evidence available for you to present.
I agree it's likely (fervently hope so, actually) that what you call "normal America" will create an enduring movement that embraces free speech over enforced left-wing correctness. However, I don't know why that movement needs to include the more toxic parts of MAGA culture like the relentless misogyny, opposition to reproductive rights, enforced Christianity, desire to ban all LGBT not just youth gender transition, love of foreign dictators and authoritarianism, making health insurance harder to get again, etc.
I have been thinking, only half in jest, that I should switch from being a Democrat to being a never-Trump Republican (moderate Eisenhower/Weld/Baker side, to be sure). If either party manages to resurrect the moderate center, that would be a movement that could endure.
I actually think that is an absolutely reasonable position to take Monica. I would quibble about things like relentless misogyny (take a week and pull up Megyn Kelly's podcasts of her radio program - I mean that seriously). You will be hard pressed to find a stronger, more able, or intelligent woman on either side of the aisle. But she also prefers her men to be men while she is being a woman. Also, it is probably worth noting that Trump did just choose a very tough and competent woman to serve as his chief of staff - something no other president has done. And lest we forget, it was he who nominated Nikki Haley to be UN Ambassador during his first term. For whatever it is worth, she was my choice for the nomination.
Reproductive rights are a difficult issue. I do believe much majority opinion is regionally based. In other words, it is entirely possible, even reasonable, that respective majorities of women would have different opinions on the issue in different parts of the country. Hence, I think the "leave it to the states to sort it out" solution is extremely wise. We conveniently forget that Justice Ruth Bader Ginsburg almost certainly would have agreed.
But it is a tricky subject. I do not think any serious person disagrees with the issues of rape, incest, or the health of the mother. But there is a reasonable question that needs to be answered. At what point does the state have a responsibility for the health and safety of the unborn child? I think a very large majority of Americans, male or female, would find an issue with unnecessarily aborting a fetus in the third trimester, much less killing an aborted child that somehow survived the procedure. I also think most pregnant women would expect the state to aggressively charge and prosecute any second party who deliberately or negligently harmed an unborn child in the womb. So how do we resolve those conflicting demands? I do not pretend to know the answer to that, but I trust the wisdom of the people - particularly where they reside - far more than the Washington bureaucracy.
Polymarket ended up cherry-picking a date to make their subscribers think they've predicted the map correctly in advance: https://twitter.com/nsokolsky/status/1854280616507834538
Everyone's been dunking on Silver Bulletin and 538 for being less accurate than prediction markets, but the per-state data shows that traditional modeling was still superior.
that's a strange criticism, since polling aggregators are of course not exogenous to prediction markets whereas the reverse can't be true without leading to feedback loops, as Nate noted in his Election Day SBSQ:
> Do you anticipate ever assigning “pollster ratings” to prediction markets and incorporating them in the model?
> No. Because I worry about recursive effects. Say our model accounts for prediction markets, so it moves toward — in this election — Trump. Then prediction markets see that and say “wow, Nate’s model is moving toward Trump!”. So Trump’s price goes up further. Then the model sees that, and Trump’s probability improves even more. And so forth.
That's a strange criticism of my criticism. Yes, prediction markets are subject to feedback from Nate's model as well as all kinds of other problems. But my point is that their accuracy isn't superior to traditional analytical methods just yet, as evidenced by looking at the sub-markets instead of just the headline market. *Why* that's the case doesn't matter, we're just discussing what numbers both models provided.
And this is why Polymarket's had to use a trick to claim they've predicted all 50 states correctly, instead of just taking the numbers from Nov 4th.
apologies, my wording was very poor. I was agreeing with you. It's the people dunking on Silver/538 that I think are levying the "strange criticism".
Nikita,
With respect, the betting models have a better track record for calling winner (90 percent or more). The models have missed the last 2 of 3. It’s not the modelers fault. Pollsters are not providing the models with worthy data.
https://polymarket.com/event/house-control-after-2024-election
Call me back when they correctly predict more than just one topline race.
I'd google Richard Roll if I were you.
Is that a Rick Roll?
"Admittedly, the sorts of Republicans who encounter our work are not a representative sample, probably being on the moderate side"
I would wager this is not correct. I am a subscriber and I am to the right of Genghis Khan. The "moderate" Republicans I know are mostly checked out and do not pay attention to news at all. They are certainly not committed enough to look at polls and models every day.
I wouldn't even call myself a Republican, but I did vote Trump, and I'm fairly moderate in that I support national health care, am concerned about climate change, support mandatory vaccines, want FOSTA/SESTA repealed, want a child tax credit, etc.
Do you mind sharing what made you vote for Trump that superseded these positions? The economy (and/or inflation?) Immigration?
Immigration, wokeness/identity politics, cancel culture, gun rights, free speech, corporate free speech, the smugness and condescension of the left, the scandal regarding hiding Biden's senility, religious freedom...
Amazing how Trump gets free ride on felonies and sexual assault and general vulgarity. And Harris, the product of public schools and broken family is dinged on free speech, smugness, condescension, and... Scandal LOL! Scandal of all things to ding Harris. Wow America. Really? You're Against scandal so vote Trump? Let other generations worry about climate change. Let the kids worry about social security. Let Trump and the brilliant RFK take away pre existing conditions protection.
The felonies are absurd overreach and will almost certainly be overturned.
I don't believe he ever committed sexual assault. The accusations are years or decades old and not particularly credible.
The vulgarity is really just a stylistic thing, not a substantive issue, but I find it refreshing personally.
"I don't believe he ever committed sexual assault. The accusations are years or decades old and not particularly credible." Forgive me if I discredit your opinion on this given that the jury of 12 of Trump's peers unanimously agreed that they believe he did when presented evidence by both Trump and the plaintiff and your lack of knowledge on the matter relatively. When one considers that your level of analysis of this matter is about on par with that of someone with median intelligence, it is extremely discouraging to consider that half of the population will have an even poorer assessment of the situation.
Hahaha he's literally on audio admitting sexual assault. Degeneracy is so refreshing! Court decisions are illegitimate! Criminality isn't a bug, it's a feature! Clearly we've got the government we deserve.
There is no "global warming." That is all nonsense designed to switch the public out of oil and coal energy and into renewables. Which I would be fine with if it's organic and not subsidized and forced, but I sure don't like the Big Lie of global warming that seems to go on forever, with never any actual evidence.
Gotcha, thank you for sharing!
wait, what do you mean by corporate free speech?
The Supreme Court ruled that people could join together to form organizations to express their views, in particular political speech, without restrictions on funding. This was a great First Amendment case called Citizens United.
The Dems are against it and have campaigned on overruling it, which is truly scary, because that would mean banning political books or movies before an election.
We have a system similar to what the Dems are advocating for in Canada (severe limits on third party spending to prevent PAC-style advertising combined with caps on donations to force campaigns to use only small dollar donations) and the result has been to essentially silence large chunks of the political spectrum in the run-up to elections. Ironically, the consensus is that it has actually harmed left-wing political parties more, because they're more dependent on fundraising/spending from civil society organizations, activist groups, and unions, as opposed to right-wing parties which do better raising moderate sums from individuals.
I would honestly argue the opposite. I’m extremely against citizens united and think that the unlimited flow of money into politics has let the rich basically buy our politicians. They don’t work for us anymore (or at least even less than they used to) they work for the highest bidder.
I’m not sure why you say this would lead to political books and movies being banned. We had campaign finance limits for over a hundred years up until 2008 and this didn’t happen.
Like Yglesias said:
https://nitter.poast.org/mattyglesias/status/1854334397157384421
I think Nate is hedging his bets on his ability to attract conservative Trump supporters, but he's underestimating his appeal. There's many Trump supporters here, often to the confusion and anger of some liberal subscribers.
He's probably assuming conservatives wouldn't want to subscribe to him because he's a liberal, but he's forgetting that we only care about merit and not identity. That's kind of our thing. He's very good at his job so here we are. I still feel that way even if the model was a little off. It's dependent on the polls that are fed into it after all. If anyone dropped the ball it was the pollsters, not Nate. Garbage in and garbage out.
Me too!
Atlas Intel beat your models. And you can read their polls free on X.
Do they publish their predicted distribution of outcomes?
What was their predicted map as of Nov 4th?
The actual map. They had Trump winning every swing state.
Can you link to it?
If they’re that good they may not be free for much longer.
They are using political poll results as advertising for their commercial services.
Their published slide decks included a half dozen or so slides about other successful results and kinds of services.
All pollsters use political polls as advertising for their commercial services. That's how the polling business model works.
I know, but it wasn't clear that the cat lady knew.
The number of Atlas advertising slides was a surprise.
Hi Nate, can you do us a huge favor, before winding down the election content...
FRENCH WHALE STORY PLEASE.
Is the 'neighbor' question not herded in the original polls, just because the neighbor question isn't in mainstream news? What's the veracity of this stuff? Has this been reported all along? Will this ever end up in the model, or is it already?
You want him to be a whale psychologist?
Apparently the whale did his own polling that asked who people thought their neighbors would vote for, and the result was overwhelmingly for Trump. 'French whale' is a total misnomer for anyone doing their own poll, and especially for someone who makes a ton of money and then walks away from the table.
https://www.yahoo.com/entertainment/seth-meyers-fully-supports-trump-141519392.html
lmao thanks
Are we supposed to be impressed after you spent - how long? - calling it a coin flip?
Jk sorta but also, you’re reaching
Polling error inside the margin is nothing to be ashamed about. I'm impressed with the correlated outcomes by state, which are less attributable to model input. I'm most interested in more analysis on the LV national polling miss and what the polling industry is planning to do to align sample size by region and close the LV/RV gap.
If you want to hear the honest truth,
most of them won't be doing anything,
Save for most likely engaging in controlled herding in the future
and *maybe* offering more cross tabs to end clients earlier.
Which seems counter productive, but this is where I always have to tell people that 'Political Polling is The World Poker Tournament for Polls, you mostly come to be seen, not to win'.
No one makes real money (and often don't even break even) doing these political polls, but you can find good regional offices and get your name up where corporate clients will see, so you do it anyway.
The poll margin, *in the regions that most pollsters specialize in*, were very good. Outside those areas, there was variance. So, in the end, they aren't going to benefit from investing.
Add in that the next election will be a completely different ball game, with no Clinton Era holdovers and Trump closing out his Lame Duck term, and you could over-correct your company into oblivion. So, best for most of them to play the safe game.
If a hundred people take a random guess some of them are going to be right but it's not necessarily more useful than having a model that quantifies the uncertainty and enumerates the possible outcomes and their relative likelihood.
Didn’t the model also give Kamala like an 80% chance of winning the popular vote?
Trump did overperform a lot on the national popular vote but that was mostly due to overperformance in deep blue states where there wasn't a lot of polling (due to them not being seen as being in play). It's not surprising that the model didn't catch that, given that its prediction of the national popular vote is mostly extrapolated from state-level polling.
As someone else put it the “model can never fail. It can only be failed”.
I wouldn't go that far - I think this certainly points to a weakness of extrapolating the state-level polling into national-level results. Maybe the imputed variance due to lack of polling should increase or decrease until the national popular vote distribution from the simulations matches that derived from the national-level polling?
And yes I understand that means she was expected to lose the popular vote 20% of the time which is not all that rare. But I suspect this was the model not working (or bad polls) and not just a fairly unlikely thing happening.
These kind of models are unfalsifiable. You can't run the same election over and over to prove the percentages. In essence, it's a giant grift.
No, a "grift" would be Nate selling a thing, while claiming the thing is something else.
This is a statistical model designed to predict the outcome of elections based on polling, and Nate has sold it as just that. It's pretty obvious to me that he has put a lot of effort and his expertise into it.
What you've paid Nate for is access to the outputs of this model, plus his commentary on it (up to and including his own punditry).
What on earth else were you expecting from your subscription?
I’m not even blaming Nate or anything but if the polls are not very good then the model is useless. Garbage in. Garbage out.
I actually think this specifically *was* an issue with Nate’s model rather than polling. Plenty of high quality national polls showed Trump tied or ahead of Harris in the final weeks. I think Nate has said that his model didn’t use national polling much, but instead extrapolated the national outcome from state polling. It seems like that would be self-reinforcing of the idea that there would be a big gap between the popular vote and electoral college outcome in favor of republicans, as was true in the last two elections, but it looks like that wasn’t really the case in the actual election outcome.
I didn't get the sense you were blaming Nate, and yeah you're absolutely right.
Honestly, I think people expect a bit too much from polls. When sampling error alone (setting aside methodological error) is 3-4pts *for each candidate*, and people expect you to predict the outcome in a state when both candidates are hovering in the 45-50 range... people are not gonna like the result.
And then I also guess with partisanship every election is going to be close. Polling misses seem to be systematic and not state by state so it just seems like a tremendous level of analysis to say polls are close but they could miss j one direction and equally likely in another. If the polls and state outcomes were less correlated perhaps the model could provide more insight and do more useful probability crunching.
Commentary on election night. Which I didn't get. Hence, I am no longer a subscriber.
They're absolutely falsifiable over the long term, just look at multiple predictions.
If you systematically looked at 100 80% predictions and the model only got them right 60% of the time, the model probably isn't very good. If it gets them right 75% to 85% of the time, it looks much better. If it gets 100% of those predictions right, the model is actually bad because it's way too under-confident.
National presidential elections are rare, but the model makes way more predictions than just the outcome of the presidential election.
Every model he runs is different. So you can’t compare some basketball model to the presidential election one. It would take literally millennium to have a data set where you can say okay there were 10 80% elections and the favorite one 8 out of 10 times. It is 100% something that cannot really be proven wrong.
You don't need 10 80% elections and 10 70% elections on so on to evaluate the model's calibration.
Instead, imagine that you can bet against the model. On each probabilistic forecast, you bet $1, and the model pays out $1/p where p is the final probability of the event you're betting on (so for an 80% model output, it pays $1.25). You add up the results of these bets, $1 on each forecast, and in the long run the model is well-calibrated if it returns $0.
You can use the other political models, though - congress, senate, governors. They all use the same basic approach as the presidential one. It's not a perfect assessment of the model, but it's pretty good. You can't really assess the model when it is saying the race is 50/50, though - Nate made that point himself a week or two ago. A model that always says it is 50/50 will come out as perfectly calibrated.
To the extent that's true, why has the actual outcome vs. his prediction always been well towards the tail, IN THE SAME DIRECTION, each and every time Trump has been on the ballot? If the model were unbiased, getting, say, a 20% tail result in the same direction three times in a row would happen ~0.8% of the time.
So now we're down to priors: Nate's model is great as is and we just need to have Trump on the ballot a bunch more times and it will all even out, OR...
"To the extent that's true, why has the actual outcome vs. his prediction always been well towards the tail, IN THE SAME DIRECTION, each and every time Trump has been on the ballot?"
I think there are several reasons for this, but the main reason is that the model is based primarily on polling, and polling has underestimated Trump consistently. You can build a model that operates on the assumption that polls will consistently underestimate a candidate, but then it's not really a polling-based model anymore.
You can make the argument that some polls should've been weighted more heavily than others, based on the results of this recent Presidential election. And that argument may now be (given the results) evidence-based, and will probably be incorporated into the next model (if there is one). But that argument wasn't based on much evidence previously.
"20% tail result in the same direction three times in a row would happen ~0.8% of the time."
This only holds if the predictions are statistically independent. I don't think that can be assumed here. Way too much going on.