Twitter | Search | |
Yaron (Ron) Minsky
Occasional OCaml programmer. Host of .
5,843
Tweets
182
Following
7,243
Followers
Tweets
Yaron (Ron) Minsky Oct 29
In other words, the last election shouldn't move you that much, but this election should move you a lot. (Obviously, depending on your priors, you may or may not end up with the view that 538 is more likely wrong than right.)
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
i.e., if we consider the possibilities of 538 was right, or actually Trump was 80% to win both times, the later becomes 8x more likely if Trump wins this time, ignoring the last outcome, and 19x more likely if you count both.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
I think it's right that it should reduce our confidence a lot. But most of that reduction comes from this time, not from last time.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
I think the math (along with some assumptions) is necessary to support your initial claim that two mistakes should move your priors. So your original claim just doesn't move me. In the absence of some math, it's hard to claim this is a Bayesian result.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
If we were real bayesians, we'd do the calculation. I think the "wrong twice" case is less of a move than you think, but if I was less rusty I would just do the calculation.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
And how you feel about it should probably depend on the details of how the event occurs.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
I think it's true that if Trump wins, you'd need to materially up the probability that there was a problem with the 538 model. But remember, 1-in-10 times, this size of upset should happen by chance, so you shouldn't be that surprised if it happens!
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
I mean, if you're thinking about politics strategically, like, where to run a get-out-the-vote campaign, then it matters. But for most of us? Sure, kind of pointless. But, then again, most of the news I read isn't exactly actionable. And yet, I keep on reading...
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
And what does it mean for them to be "equally wrong" this time? Are you talking about just the up-or-down question of who wins? Are you thinking about the state-by-state breakdown? About the final poll results in each state, or overall?
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
Yeah, still don't get it. Part of the issue is that "wrong" is a subtle thing. If someone makes a probabilistic guess, what makes it good or bad? The question is not whether it works out, but whether it was a good estimate given the available information.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
At least, about as good as they've ever been. The idea that polls don't work anymore is basically unsupported by the evidence.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
But as they point out over and over: the polls are actually pretty good!
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
The "lol, probabilities don't say anything" view just doesn't add much to our intellectual understanding of how one makes predictions.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
i.e., if you take the problem seriously, this is kind of the only approach. And you can evaluate 538 in a reasonable way by looking over multiple predictions over longer time periods.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
Sure, but it's a weird thing to complain about. The only reasonable way to phrase projections is probabilistically, and you only learn thing about probabilistic projections incrementally and probabilistically as data comes in.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
538 isn't perfect, but they're not a bunch of hacks.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
And no, 538 did not have Biden at the same levels as Hillary, at this time in advance of the election. The day of the election, they had Trump at 30%. This time around, Trump was at 30% a couple months (?) ago, and it's not down to 10%. Those are very different.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
Huh? If you predict something is 30% to happen, and it happens, are you wrong? Honestly, I can't tell if you're asking a serious question or just kidding around.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
Replying to @jordwalke
But, that's just not true. 538 has a 10% chance for Trump to win. That's quite a lot of uncertainty.
Reply Retweet Like
Yaron (Ron) Minsky Oct 29
Replying to @yminsky
And the knowledge is gnarly and complicated, rather than being some clear idea someone had about how to design a piece of software. Hence, the extraction is fundamentally time-intensive.
Reply Retweet Like