Model fits in a tweet and predicts all elections since 1960 | Hacker News Reader

Model fits in a tweet and predicts all elections since 1960

twitter.com

52 pointsrahimiali58 comments

For those missing the context, this model was a reductio ad absurdum response to https://twitter.com/MartinDaubney/status/1314590150245183488

You can overfit any small dataset using a relatively simple error reduction method.

My model is a hashmap that contains my training data. Missing keys are forecast uniformly randomly. It perfectly matches 100% of my training data.

Exactly. The troubling part is that even with overfitting his model still missed two elections.

That’s probably intentional, no one gets 100% on the exam.

So... the bug is a feature?

As Nate Silver complains about often, there just aren't that many presidential elections since 1960, it's way to easy to build some overfitted model that seems accurate. As another comment mentions, we could all just build a nth-order polynomial to achieve a 100% accurate prediction on previous data.

I could make an nth-order polynomial that selects the correct outcome of the past n elections ... and if I shroud it in enough mystery and red arrows I bet at least a few "news" sites would share it

And with such a high order, the predicted next result would probably be crazy, like -1000 on a scale of 1-10 or something

Haha that's not even a problem: you can map the reals to (0, 1), so just do that for whatever number that the model spits out and then multiply by 10. Bam, there's your model that perfectly predicts all of the past presidential elections.

"With y=0 for Joe Biden, and y=1 for Donald Trump, my model predicts y=154476802108746166441951315019919837485664325669565431700026634898253202035277999"

You just need to then overlay a function that takes the remainder of the prediction / 2. Perfect model fit and a seemingly sensible prediction for next time :P All ready for media consumption :)

/data scientist here people. Please don't actually do that.

Please do! ^_^

Just make sure you have a model prepared indicating a win for whichever direction the particular news agency leans. CNN gets the Biden-wins model, Foxnews gets the Trump wins model. Democracy Now! gets the Bernie-wins-outta-nowhere model.

Funny but disingenuous, since "predicts" implies that the model was created before the 1960 election. This no more predicts a 15-bit value than getting two specific bytes out of a PRNG with a specific seed.

It's just like that post a few years ago where newspapers used various dodgy "predictors" like "a [state name] president has never lost a reelection", and someone had collected a bunch of these which turned out wrong.

This is in response to a model that similarly "predicated" presidential elections (using the algorithmic power of overfitting)

You might be thinking of xkcd 1122 "Electoral Precedent": https://xkcd.com/1122/

Haha, very funny. It's based on the first three characters of each name.

Hmm... I don't see it. Can you explain further?

    def win(A,B):
        s = ((50*(ord(A[0])-ord(B[0])) + 114*(ord(A[1]) ord(B[1])) + (ord(A[2])-ord(B[2]))) % 199) % 2
        return B if s else A

ord() basically casts a char into the corresponding integer (ord("A") == 65, for instance). The whole algo is basically just some math based on the first 3 characters in each of A and B

Numerology has been a time-honored way of making predictions for millennia.

For the longest time didn't they always talk about how the tallest candidate had historically won since some date... And then it stopped working.

Although... if I was building an election model, I would include candidate height as one of my predictors.

Not using ASCII letter values, though.

Funny how 'Jimmy Carter' wins against 'Gerald Ford', while 'James Earl Carter' loses.

it's python and saying A[0] means first character and ord returns an integer for the current unicode glyph (like char, but for unicode) i.e. the seed are basically the first three characters.

I am reminded of the Atari 2600 game Pitfall and its use of a polynomial counter to procedurally generate the game world[1].

[1] https://www.youtube.com/watch?v=MBT1OK6VAIU&feature=youtu.be...

Also see Elite [1] for a similar procedurally generated game universe.

[1] https://en.wikipedia.org/wiki/Elite_(video_game)#Development

Given how unique casting votes will be this election due to the pandemic, I don't know how anyone can be comfortable making predictions about the outcome based on past data.

Polling depends on having a good screen for Likely Voters, which is difficult in a normal year but this year is going to take a miracle to get correct. Given that pretty much every state is winging it a slightly different way, coming up with models that work for even just the battleground states is going to involve as much luck as skill. Pollsters that get it wrong will end up with a credibility drop more than deserved while those who get it right will get more of an increase than deserved.

Did you look at the tweet and it's content?

Nate Silver and pretty much every pundit got it spectacularly wrong in 2016 and had zero hit to credibility, he and his colleagues are still somehow trusted.

Same story across the pond for Brexit, which polls botched spectacularly.

2020 is too uncertain to be predicted in my opinion, there's way too many things we haven't seen before.

Nate actually gave Trump a reasonable chance of winning (I want to say 20%?) which if you understand how probability works doesn't really hurt his credibility. 20% chances happen all the time.

The people who really took the hit were orgs like NYT who gave Hillary like a 99% win probability on the day before the election.

29% chance, actually. Washington Post had a piece that mentioned this today.

"On Election Day that year, FiveThirtyEight’s forecast gave Trump a 29 percent chance of winning. This was a better likelihood than was reflected for Trump in a number of other forecasts, though it still amounted to just under a 1-in-3 chance he would win. The site’s Nate Silver wrote that there was a 1-in-10 chance that Trump would lose the popular vote and win an electoral college majority, which is of course what happened."

https://www.washingtonpost.com/politics/2020/10/12/how-think...

The contrast between the way sport win loss percentages are dealt with, where 10% chances happen 1/10 times so that you get used to it and the way that people take politics predictions is dramatic.

No doubt politics is harder, but people need to get calmer about these things. Prediction with 70% confidence and the other side winning is not wrong.

Right before the election he actually called it a much closer race than most polls suggested. https://abcnews.go.com/Politics/nate-silver-predicts-close-2...

Number wise it 28.6% Trump wins which is an edge, but not a huge one. https://projects.fivethirtyeight.com/2016-election-forecast/

The real issue with these forecasts is so few elections happen that you can’t actually tell much about how accurate the predictions are.

Trump did not have a 28.6% chance of winning at all. The number was simply false because the underlying polls were fundamentally skewed in multiple states.

Nobody trusts polls as completely accurate, predictions are always fudged to take into account such things.

Now, Trump had a very healthy margin in the electoral college which might make it seem like an easy prediction, but he only hit 46.1% of the popular vote. Hillary on the other hand only hit 48.2% which shows 5.7% of the population was willing to go 3rd party for whatever reason.

In the end it’s never a single issue, Hillary supporters stayed home on Election Day in part because she was predicated to win. That’s the kind of issue that makes predictions so difficult. It’s rarely a question of undecided voters, but rather who is willing to show up to vote. Something as simple as a little rain can swing elections.

That's not supported by the data. Hillary was significantly less popular among black voters than Obama. In fact, if black voters had an identical turnout in 2016 than in 2012 Trump would've lost.

It's easy to see why given her history, too. She has a history of pretty racist remarks and her husband spearheaded the infamously racist Crime Bill.

What’s not supported? Those are the actual results of the election. Also, 2012 and 2008 elections both had under 2% voting 3rd party which is more normal.

But he did. Specifically according to Nate silver on Election Day in 2016.

The state polls were wrong by large margins, therefore Nate Silver's prediction was built upon a mountain of false information.

Some of the polls were wrong by margins larger than the purported margin of error, implying the entire poll was structured incorrectly.

Those same polls have not re-adjusted since 2016.

> Those same polls have not re-adjusted since 2016

This is incorrect. Almost everything you've said here has been incorrect.

https://fivethirtyeight.com/features/what-pollsters-have-cha...

This seems to be a misunderstanding of polling, of models built on polling, and maybe on statistics in general.

There was some polling error in 2016, it's true, in several states that turned out to be determinative. Uncertainty can be and is built into models based on the quality and frequency of state polls, and 538's model certainly takes the polling errors of 2016 into account by adding more uncertainty to states and pollsters which got 2016 wrong.

That said, most of the shifts in state results were within the margin of polling error, which is why Silver's model gave Trump a much higher chance of winning than any other major models did.

For 2020, the current probability is 14%

https://projects.fivethirtyeight.com/2020-election-forecast/

Reminds me of 2013 IOCCC entry: https://www.ioccc.org/2013/cable1/hint.html

Serious question: how does one go about finding such functions?

I mean, given N inputs and known outputs, how does one design a function that is less than a N-order polynomial? Are there techniques or tools available, besides the elementary examples of simplifying a Karnaugh map?

This looks and smells like it was evolved by a genetic algorithm to produce those (and perhaps only those) results.

It's based on the characters in the candidate's name, so yeah (though I doubt a genetic algorithm was involved)

That has got to be the best compression of a deep neural network that I've ever seen. ;-)

The training set equals the validation set?

All USA elections.

(There's a certain predictably about gratuitously grandiose claims actually being constrained to 5% of the world.)

Analogous to this relevant XKCD from 2012: https://xkcd.com/1122/

For anyone wondering this algorithm indicates a Joe Biden victory.

He tuned it to ensure that outcome.

More troubling results for 2024 however:

win("Joe Biden", "Gary Busey") == "Gary Busey"

Only if Vincent Pastore is his running mate.

Unfortunately, Biden is significantly shorter than Trump so it isn't a safe bet he'll win: https://en.wikipedia.org/wiki/Heights_of_presidents_and_pres...

Significantly? Isn't Joe Biden 6'0"? Trump is maybe 6'1" (I know he's claimed 6'3", but since he's actually shorter than Justin Trudeau....)

https://www.macleans.ca/news/world/the-g7-group-shot-where-d...

I think the source on trump being as tall as he says he is, is from his own doctor (remember the kerfuffle over that, simpler times!)

Reply on news.ycombinator.com