A Call to Adventure | Hacker News Reader

A Call to Adventure

overcomingbias.com

71 pointsr72110 comments

Hanson's "Adventure #1" is the promotion of futarchy, or rule by (prediction) markets. This is one call I fervently hope goes unanswered. It's the most dangerous sort of bad idea - namely, one that produces good results in small-scale trials.

Prediction markets perform great, so long as they don't matter. The idea is that you make a contract that pays out if some future event occurs (eg "pay the holder $1 if Donald Trump is re-elected in 2020"), and then you trade it on an exchange (like the stock market). If the market price is $0.50, that means a 50% chance of Trump winning. This produces much better estimates than listening to pundits on TV, because traders profit by being right (rather than, eg, getting good ratings, or looking good to their boss). Prediction markets perform very well in studies, and so Hanson et al propose we should use them to make all our decisions.

That's where things go from empirically great to predictably terrible. Goodhart's Law[1] says that if you use a measure to control a process, it stops being a good measure. (Think of the Soviet nail factory which was rewarded for each ton of nails produced - so produced a single, massive nail each day.)

If you make decisions based on a prediction market, any successful pump-and-dump scammer can control your decisionmaking. Put another way, every decision can be bribed (by someone willing to lose enough money moving the market price).

Proponents counter that the market will self-correct (eg it's difficult to move IBM's stock price much just by betting on it, cause lots of people will bet against you). But the S&P 500 is a well-developed, liquid market. If futarchy were ever implemented, the policy equivalent of the Pink Sheets[2] would be carnage.

[1] https://en.wikipedia.org/wiki/Goodhart%27s_law

[2] http://www.investopedia.com/ask/answers/201.asp

Maybe the way to solve this bug is with a contributor limit like we have for campaign funding?

Maybe, but trying to make a market better by regulating it sounds wrong.

There are other kinds of prediction systems that don't use markets. One researcher developed a system that just recruited a bunch of random people and took the predictions of those with the best historical track record. And found it was much better than chance or predictions by CIA analysts with a lot more information and knowledge.

Others have found that markets can work almost as well with just play money. It's just a game with no real stakes, but the wisdom of crowds and competitive human nature still makes it work. It's much much harder to manipulate though, because you can't spend real money to buy predictions. You can only get play money by organically earning it over a long period of time making actually good predictions.

I don't think real prediction markets are as bad as you think though. Especially if all bets must be public. Then manipulation and conflicts of interest are easily detectable. And you can earn money by betting against them. People would quickly make bots that just bets against new accounts betting lots of money, or inverts the trades of people with historically worse than chance track records. People will develop meta models that make bets based on the trades of other predictors. And come up with crazy machine learning systems to detect manipulation.

If that somehow doesn't work, the best system might be a combination. E.g. have a group of predictors who make the final decision, but have access to the results of a prediction market. And can decide when to trust it and when not to.

Of course that shouldn't be necessary. Because if there is anyone that can tell better than chance when the market is wrong, they could just make tons of money betting on the market itself. And in the process correct the error.

The Good Judgement Project (for that was its name) is pretty awesome - but if you mechanically connected it to the policy apparatus, I'd expect Mr Goodhart on the next train.

You've proposed a few possible countermeasures (play money, publication of bets, a guess at the equilibrium outcome of a vastly complex HFT game). Each of these might work, but (a) isn't actually what Robin Hanson is proposing (b) hasn't been tested as a defence against highly motivated adversarial actors.

And they would be motivated. The threat model here is "manipulate this market and directly influence policy in your favour". They say "money in politics is like water - it finds all the cracks", and politics at least has humans in the loop. The track record of people betting against it is...not promising.

Please show me how any of the ideas I've proposed can be exploited. I guess in theory all the people involved could be bribed - but that's already true in our current system! At least this distributes the process over many more people and vastly increases the number of people you need to bribe.

Let's try something and see what happens. It doesn't need to start with the government, even private businesses implementing these systems would be a great test bed. Then we can work out the flaws, if there are any. The potential benefits of improving our prediction making ability is enormous. This is a problem that, in an ideal world, our society would be devoting tons of resources to solve. The status quo sucks and we shouldn't just give up and accept it.

meredydd's specifically voicing concern with decision markets, not prediction markets. Hanson's Futarchy paper outlines the difference (alas, wikipedia unhelpfully treats them as synonyms).

Hanson's "Adventure 1" is quite a bit more specific than "improving our prediction making ability" (which of course could unlock tremendous value).

Considering the fact that Hanson knows all of this, and still thinks his idea is sound, what would be his response?

As many have long expressed this concern, I and others have long done research into "manipulation" of prediction markets. We've done theory, lab experiments, and field tests. We find prediction markets are especially robust in resisting such influence, compared to other institutions and mechanisms.

I've been working on the second one for a while on my website, here's an example: https://psychology.wtf/social_coherence.php . However, that's just for fun, it's not a final answer to anything. Correlation correlated with other correlations.

Reply on news.ycombinator.com