Pablo_Stafforini

Comments

AGI Predictions

The probability I would assign to #8 intuitively is about 0,41. Math based on my other three predictions yields (doing the calculation now) 0.476. I am going to predict the math output rather than my intuition.

I think the correct response to this realization is not to revise your final answer so as to make it consistent with the first three. It is to revise all four answers so that they are maximally intuitive, subject to the constraint that they be jointly consistent. Which answer comes last is just an artifact of the order of presentation, so it isn't a rational basis for privileging some answers over others.

2020 Election: Prediction Markets versus Polling/Modeling Assessment and Postmortem

The contracts are denominated in USD, and they pay in that currency. But you trade on margin, and the collateral can be in any currency (crypto or fiat). In your example, you get back the BTC plus 15% of what that BTC was worth in USD when you made the trade. 

Incidentally, TRUMPFEB is now trading at 0.16 (i.e. implied 16% chance that Trump is president next February). This looks insane to me (and I have bet accordingly). I'd be curious if you or others have further thoughts on what might be going on.

2020 Election: Prediction Markets versus Polling/Modeling Assessment and Postmortem

I'm not sure I understand your argument, given that FTX allows traders to keep balances in both USD and BTC, but in any case historically FTX prices have been in line with Betfair/PredictIt prices, so I doubt this consideration is relevant.

Bet On Biden

I'm too lazy to look it up, but I did research this a couple of weeks ago and found that 538 had indeed outperformed the markets both in 2008 and 2016 (I wasn't able to find data for 2012). This is not very informative, though, since it's just a handful of cases. Much better is to look at the state-level predictions and use brier scores as a measure of forecasting performance.

Bet On Biden

man do I wish that there was just one large market with minimal fees

Such a market exists, though unfortunately it is restricted to countries where most LW users are not citizens of.

Bet On Biden

Only the first article in the comment is by Silver, on whose expertise the original poster is basing his recommendation. That article doesn't discuss mail-in ballots or voter suppression, and in fact his main point is that the time remaining until election day (almost three months when the article was written) combined with uncertainties due to Covid-19 meant that the race was still open back then. Those considerations have much more limited force at present, when only 16 days remain, and Biden's lead has widened considerably. 

If you've been at all listening to Silver recently, you'll know that he thinks his model probably underestimates Biden's chances. This shouldn't be surprising, since as Silver acknowledges, in this new version of the model he has made a special effort to build conservative assumptions into it.

In any case, I would encourage people hesitant to bet for Biden to resist the temptation of "throwing in a bunch of considerations" for why the models may be wrong, and instead try to calculate what the correct forecast should be in light of those considerations. For example, if you think mail-in ballots will be a big factor, try to estimate the magnitude of this effect.

Following my own advice, I just built a simple Guesstimate model of the impact of mail voting on the popular vote. I created the model very quickly, so if anyone spots any errors, please mention them below. And if you think some of the parameters should be different, simply copy the model and adjust those parameters to your satisfaction. Note that the effect of "rejected" in-person ballots is not modeled. This effect favors Biden, since a greater proportion of Trump votes will be in person, and hence susceptible to being "rejected" (i.e., not cast due to failure to bring an ID, long lines, inability to find a polling station, etc).

ETA: The upshot of the model is that mail voting shrinks the expect popular vote gap between Biden and Trump by about 2%. If we assume that the electoral college gives Trump a ~2% popular vote advantage, the model implies a drop in Biden's chances of winning the election from 87% to about 79%. [I modified the model and improved some of the estimates, and now the effect is less than 1%.]

(Disclosure: I have bet a total of USD 12k on Biden, mostly back when his odds where roughly equal with Trump's.)

The Alignment Problem: Machine Learning and Human Values

The book is dedicated "for Peter, who convinced me". Maybe that mysterious Peter is the ultimate cause of Christian's interest in Al alignment and his decision to write a book about it?

Forecasting Newsletter: September 2020.

a frustratingly well-paywalled, yet exhaustive, complete and informative overview of the IARPA's FOCUS tournament

Since you quote from a section that is behind the paywall, I assume you have access to the article. If so, could you make it available? Or just send it to me (name@surname.com) and I'll upload it to my site and post a link to it here and on LW. Thanks!

Load More