Nobody is going to care about any of his lisping or details about how akshually i gave him a 13% chance after the fact if Trump wins, given Hillary. He is all in and he knows it.

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41198675)

Hint: multiply 30% x 13%.

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41198812)

A Trump win is not a "15% event happening" , it's not like oh wow we drew a jack of spades what are the odds! Or that it's some stochastic mechanism like a 'random' drawing of a lottery ball, and a low chance event just occurred

Narrating as such to his readers is betraying his own favored philosophy for analysis - bayesian - where it represents a degree of belief in an event occurring, and you update those beliefs by incorporating new data/ info into your set of information

A better analogy would be his model is a measurement device, and it's too low resolution

A Trump victory occurring just puts reality into higher resolution for everyone to see

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199724)

If you lived in a world with access to perfect information, you could probably build some sort of model to accurately predict the election with 100% confidence. But polls and all of the information we have are imperfect, so in the real world you can only ever come up with a probability. I’ve only ever argued that Nate Silver’s whole thing is pointless because it’s not falsifiable.

But theoretically, if you somehow did come up with a model that did accurately came up with the correct odds based on all available information and it came out to trump having a 6% chance of winning two terms, then Trump actually winning two terms wouldn’t prove the model was wrong. The Bears won 10 coin tosses in a row last season—-unlikely shit happens.

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199069)

If you bought a powerball ticket and I told you that you had a 99.99999% chance of losing. If you ended up winning, would that mean that my calculation was wrong?

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199137)

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199465)

Show your math here. This is a completely baseless assumption otherwise.

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199637)

The validation he provides on his "predictions" aren't actually predictions at all - if you look at what he puts out - it will be like "out of events our model predicted would come true with at least 0.8 probability, 87% of the time these events actually did come true"

This is very deceptive presenting this as validation - another way to translate what I just said was "out of events that were overwhelmingly obvious they would come true, 87% of the time they actually came true"

He's only validating a high threshold of precision which actually excludes a huge set of events in that field

What it doesn't tell you is "out of all events that actually came true, what % of these events did our model *predict* would come true"?

The reason he doesn't give this is why Taleb / anyone w/ statistical training and balls enough to speak out publicly against him (small set of ppl in the world) have criticized him relentlessly - b/c all he's doing is what i said originally , he's not actually willing to make *predictions* by giving a decision boundary on the model in the form of a threshold for the probability , implicitly 0.5 but by no means does it have to be 0.5, by which to say above this threshold the event is predicted yes/true, below is predicted no/false - doing this would allow actual validation as to the quality of his models, and it's not surprising the article we saw the other day showing using his model loses 6% over going w/ long-odds in soccer matches using his soccer predictions - ie one of the areas he CAN be validated in is his soccer model b/c it outputs expected goals, and it gives you a return of -6% against long-odds from that site that wrote the article

What he's doing when he puts out "validation" is he's cranking up the threshold to 0.9, restricting the validation set to only events that are easy to predict instead of all events, and saying "every time our model predicted the Dodgers would beat the Pirates, the dodgers won 94% of those games"

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199860)

Yeah that's more or less what's happening - and as you point out the degrees of freedom restriction that small ends up w/ some weird stuff - but it's all ultimately proprietary so nobody really knows , but what's especially puzzling b/c one of the largest benefits of Bayesian models are the transparency and flexibility to specify variance components , and incorporate outside information into the model

if you read through this and skim the comments especially, he definitely has some explaining to do when the basic mechanics of his model work like this....

As has been pointed out - he ends up with situations like Trump winning CA but only having 44% chance of winning the election , similarly w/ Biden taking alabama, or biden having greater chance of winning Alaska than trump of HOLDING MI, or Washington state being negatively correlated w/ Alabama

Or trump winning new jersey makes his chance of winning Alaska 58%, wtf?

The article I linked is a 100% must read IMO

What jumps out to me is Gelman showing how low the between state correlations are, translating to some weird ass tail behavior like trump winning CA but losing election

There's all kinds of weird artifacts - Gelman is being polite and asking "whether i had an error in my code" - you have to read between the lines of this article in our current environment of lib groupthink, this article translates to something more like a bomb than a polite inquiry once you adjust for that

Most suspiciously - Gelman brings up how low the between state correlations are, often being negative! weird tail behavior, and too weak of correlation even in the center - why is this significant? Well it's suspect as hell for the exact reason we're in right now - it's becoming more and more clear that Trump's chances of winning FL are way higher than anybody thought , and if we were to be told right now Trump wins FL, I think most of us would place Trump at > 0.5 to win the election..... So it's suspicious that in a year where we might have information like this on a state like FL b/c of expanded EV - that his model is behaving like this, that his statements to gelman make little sense, etc, it's suspicious b/c his model doesn't have to move any even if say FL looks DISASTROUS for Biden.

What it looks like to me in the media / pollsters etc - is that all bets are off this year - and they'll risk anything to suppress counter-narratives so they can basically suppress turnout on the right making ppl think it's over - which obvious won't work b/c they don't monopolize media modes like they used to.

But the article and comments are a must read IMO

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41200129)

you still don't get it man

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41198816)

one of the fascinating aspects of American punditry is that you can be wrong again and again and AGAIN and it doesn't matter. witness the Lincoln Project faggots all of whom, in a just world, would be in federal prison. Instead they are milking out million$

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41198849)

I know he's a climate change skeptic (insofar as libs blame EVERYTHING from hurricanes, floods, and fires on "climate change," with no evidence of causation). Where else does he stray from liberal orthodoxy?

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199178)

If he really sticks to his guns / polls hardcore, I think you're right, mostly b/c his main purpose being served right now is *therapeutic*, not analytic. Take a trip through his website comments or twitter comments - it's like a cult, anybody trying to debate him in there they literally regurgitate his talking points, and you can tell they're heavily coping. Of course we all do that to some degree - but the level of certainty and how much they lionize him is obvious that they're needing to believe he's right to extremes way beyond what any of us would need to.

Anything can happen, but he can just as easily walk back the predictions, blame bad polling, explain it as "low probability event just like i said was possible, etc" - part of me thinks this is possible, part thinks *no fucking way* he gets away w/ that this time given how furious ppl would be w/ him for false belief if Trump ends up winning

It's hard to tell what'll happen - our arguments about polling aside, we also shouldn't underestimate the gatekeeping effect. Part of what they talk about on the PPD webcast is that despite having a clear track record of outperforming literally every single pollster out there, aggregators, left-leaning polls, right-leaning polls etc - the polling 'establishment' refuses to include PPD, or even engage with them / return their calls. He's definitely a Trumpmo - but in 2018 he predicted DeSantis and gosh what was the other upset, while at the same time telling Republicans they're overestimating their house performances and he wasn't getting the results ras / trafalgar etc were getting in the house - even sounding the alarm early in the night when it looked like Dems would massively underperform all around he said no by the early results he could tell.

I think his audience will blow up after being right and he'll get more funding (he's funded by small donations of followers) - but Baris will not get the official creds needed to be considered a "serious" pollster by Silver/Wasserman/Cohen etc, regardless of how well he performs....

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41200002)

so true.

but i some point i concluded that being wrong over and over is the price of admission to the elite-insider game. it's how to show that you will do the necessary work.

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199692)

care to elaborate?

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199698)

Barnes is on it. Silver knows about many of the frauds committed by the mainstream media and univ polling outfits.

He needs to be sued.

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199693)

https://pbs.twimg.com/media/ElVyS25WoAA6G2Z?format=jpg&name=small

https://www.gstatic.com/tv/thumb/persons/685771/685771_v9_aa.jpg

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199705)

The validation he provides on his "predictions" aren't actually predictions at all - if you look at what he puts out - it will be like "out of events our model predicted would come true with at least 0.8 probability, 87% of the time these events actually did come true"

This is very deceptive presenting this as validation - another way to translate what I just said was "out of events that were overwhelmingly obvious they would come true, 87% of the time they actually came true"

He's only validating a high threshold of precision which actually excludes a huge set of events in that field

What it doesn't tell you is "out of all events that actually came true, what % of these events did our model *predict* would come true"?

The reason he doesn't give this is why Taleb / anyone w/ statistical training and balls enough to speak out publicly against him (small set of ppl in the world) have criticized him relentlessly - b/c all he's doing is what i said originally , he's not actually willing to make *predictions* by giving a decision boundary on the model in the form of a threshold for the probability , implicitly 0.5 but by no means does it have to be 0.5, by which to say above this threshold the event is predicted yes/true, below is predicted no/false - doing this would allow actual validation as to the quality of his models, and it's not surprising the article we saw the other day showing using his model loses 6% over going w/ long-odds in soccer matches using his soccer predictions - ie one of the areas he CAN be validated in is his soccer model b/c it outputs expected goals, and it gives you a return of -6% against long-odds from that site that wrote the article

What he's doing when he puts out "validation" is he's cranking up the threshold to 0.9, restricting the validation set to only events that are easy to predict instead of all events, and saying "every time our model predicted the Dodgers would beat the Pirates, the dodgers won 94% of those games"

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41199869)

So there's no one correct answer i can give you

But it's not so much that you shouldn't take his 12/100 as a probability , it's how you interpret what a probability means is the philosophical question

Classical statistics would say a probability is the limit of an infinite sum of bernoulli trials whether an event occurs or not, ie a probability represents sampling variation - the number of times an event would occur if we repeated the experimental conditions enough times

A bayesian would say that there's just the event, and a probability represents our *degree of belief* in the event occurring

These days, there's not a neat philosophical distinction between these two as different ideologies , rather they're just thought of as different tools for a job

Taleb's point is that probabilities for such an event like this should only be interpreted as odds, ie degrees of belief you're willing to actually bet against - which has the degree of psychology of betting b/c once you have "skin in the game" it lessens your motivated reasoning so that you're more likely to seek out the truth when you bear the costs of being wrong

I think the issue with Silver particularly isn't so much how we interpret his probabilities , it's that

a) that probability is only as good as the underlying mechanics giving rise to those simulations , which is more to Taleb's point

b) by refusing to place a decision boundary on his model, eg saying "above 0.6 means I predict yes, below 0.6 means i predict no" , we can't ever validate the quality of his model properly, and he only releases validation against highly likely events

His hiding the ball is perfectly natural for anyone in his position , you'll see the same thing done by others it's very common, just not from eg Gelman / Taleb for example

By not giving us a decision boundary, of course we'll naturally assume >0.5 means predict yes, but that is not necessarily the case at all in many domains - for example in credit card application fraud the tradeoff between false positives and false negatives will be much different, and so you'll tune your boundary threshold accordingly since being wrong in the direction of letting a fraudulent application get through is much more expensive than accidentally rejecting a legit application

But the heart of the issue IMO is Taleb's comments, what i listed in a) and b), and

https://statmodeling.stat.columbia.edu/2020/10/24/reverse-engineering-the-problematic-tail-behavior-of-the-fivethirtyeight-presidential-election-forecast/

might end up being some really damning evidence of misdeeds if trump wins.....

(http://www.autoadmit.com/thread.php?thread_id=4663464&forum_id=2#41200272)