I was waiting impatiently for the active inference punchline, which I think might be rather profound, insofar as you've just illustrated that markets are virtual organisms.
Thanks for the writeup, Scott! I've honestly been super thrilled to see all the different use cases our community has figured out for the prediction market mechanism; I think it really validates the approach of letting people create markets on anything they want!
One more example of a bounty/action market that worked pretty well: Elizabeth & I posted a $100 bounty for some data aggregation work for Slime Mold Time Mold's hypothesis of contamination-led obesity. It got two separate takers: https://manifold.markets/Austin/100-bounty-will-someone-combine-the
Would love to figure out how to better support this natively -- if anyone has suggestions, please reach out!
(1) "What will I (Scott Alexander) rate as most promising when I do a deep dive into the research on pregnancy interventions?"
I'm going for the "Natural birth versus Caesarian"
(2) "They’re betting on whether someone will clean up the (currently disastrous) backyard. They’ve said pretty openly that they’re hoping someone will buy a lot of “yes”, take care of the backyard project, and then take all their money. "
Could they not just pool the money and hire a landscaper or groundskeeper? When my back garden was a disaster, after a long time dithering about "something needs to be done", we finally hired a guy to come cut the hedges, cut down and take away the tree that was growing in the wrong place, and do a ton of weeding and spraying. Result! If you're waiting for someone else to do it, it will never get done - "what's everybody's business is nobody's business".
I'm not sure how the action market helps with the doctor situation. Isn't creating an action market for it just worse than the solution of asking a doctor to help, paying him if it does and demanding that he pay you if it does not? Your complaint with that system seemed to be that a risk-adverse doctor who was less than 100% confident might pass, but it seems like the action market would have the same issue here.
With respect to the Scott Aaronson thing, my understanding is that the American Math Association has come up with a solution. Or maybe some other math organization; it was explained to me a couple years ago, so one or two details may be off.
Most people who submit proofs for the Riemmann Hypothesis (or any other similarly famous problem) are are not so correct that a PhD student can't take apart their argument. So, the AMA charges about $100 per page of the proof, and then gives it to a PhD student to dissect. The student writes up a clear report on why the proof doesn't work, which is then returned to the author, and the student gets some cut of the money to use as travel funds.
Now, this may not still be totally accurate, or perhaps the person once telling me it was not accurate in their description, but it still feels like a good system.
>I once joked that instead of lower courts, we should have prediction markets on what the Supreme Court would think of a given case. Same idea, higher stakes.
Related idea I've had whenever there's a SCOTUS vacancy: a major qualification for appellate court judges to be elevated should be the extent to which they have exercised their delegated authority faithfully, which is to say, when cases they've decided subsequently make it before the highest court, do the Justices agree with them? 5-4 splits would have the least weight either way, 9-0 agreeing/disagreeing with the judge would have the greatest favorable/unfavorable weight; more emphasis given to cases in which the judge was in the dissent at the Circuit.
Isn't the main difference for incentivizing actions via the stock market vs. prediction markets is that stock markets are highly regulated and large transactions such as placing a major short on Tesla the day before Elon Musk is assassinated would be easily tracked by the SEC/law enforcement. The companies running prediction markets could presumably track these types of things and inform law enforcement about them, but would they necessarily have the infrastructure/processes in place to do so? And would they all voluntarily do that? Surely some privacy libertarian style prediction market would advertise how it doesn't track any user data or supply it to the govt. And what about blockchain prediction markets using anonymous untraceable cryptocurrencies like Monero?
Most other major markets where you could make large sums of money by taking a huge position and purposely affecting change in that market has some form of regulatory overview that presumably helps deter these kinds of things.
Note that, while it's obviously not for money, fantasyscotus.net is quite real and could quite easily be monetized in a different regulatory environment.
For many of these applications, I think one of the biggest challenges for prediction markets will be to beat the crude approximations to prediction markets that we already self-organize on an ad-hoc basis. To take an example from this post where I have some firsthand knowledge, I indeed don't read most of the P=NP proofs, refutations of quantum computing, and other revolutionary developments that cross my field of attention -- but that can be overridden by colleagues I trust staking some of their reputation that, hey, this is interesting, this is worth looking at, etc. Which means that a prediction market wouldn't just have to beat nothing whatsoever, it would have to beat THAT.
In the meantime, though, if it's really important to optimize my use of time, then there's surely lower-hanging fruit to be had, like me not spending hours per day laying in bed with my phone scrolling through social media like now... :-)
The phrase "the market is wrong" makes no sense to me.
Consider a market created by punters on a horse race. If the longshot going off at 40-1 odds wins, it does not mean the market was wrong. That horse *should* win about 1/40 races. OTOH, if the longshot wins 40 races in a row, that market probably isn't functioning. More important to note, though, is that if the favorite wins 40 races in a row, that also means the market isn't functioning.
If you want to know how well a market is functioning, give the market a Brier score across at least dozens of predictions. The EMH is true by tautology if you have a true market.
This sentence has a minor error where the order of 2 words is reversed, but it happens to be an error that tricked my internal parser into going down a garden path that actually took me a nontrivial amount of time to back out of:
"I throw these out all unread the time when I get them in the mail"
I considered doing the fix Valinor's backyard myself thing, but (a) it was unclear to me that they were actually angling for that and I thought it might be intrusive if I just showed up and started touching their stuff, and (b) if I'm going to do manual labor I want to be paid in real actual money and not Monopoly money ¯\_(ツ)_/¯
On why we don't have more geniuses: "His summer following his father around learning medicine was probably good for him, but not outside the bounds of what still happens today (I followed my father around learning medicine)."
On whether small prizes are great for thoughtful, complex: "I put in 20 hours of research and won. If I’m typical, this is proof of concept that [...]"
This is maybe not the most effective argument that parental tutoring is less than very effective, or that medical students can be cheaply incentivized to do good lit reviews.
In regard to the historical prediction market, I wonder how you could use them to judge the consensus on past events. There are many things that happened in the past that are still unclear, with papers being published explaining that this and this happened and two months later a counter-paper that says it totally did not. It is often hard for non-experts to
1/ judge what is the most plausible explanation of what happened,
It will be probably be too obscure, too niche, too complex and too irrelevant for it to gain traction (let me know if you got better ideas for a disputed event, which should be easy to find). I thought of using the same technique in my previous experimental market, as in "Convince me that XXX is true and I'll resolve the market this way" (https://manifold.markets/TaoSumer/are-there-more-doors-than-wheels-in), but I feel like such a world-changing event is too complex to ask people to debate constructively in the comments. I guess the best way to make it interesting would be to invite historians who wrote papers on Stalin deaths to bet money and debate/argue, and then get the crowd to decide who got the best arguments - tho rethoric would probably come in the way, with people voting for the best speaker/writer.
Do you have any ideas on how we can make such markets better? I can implement them for the Stalin market - or create a more interesting one.
Oh man, you wrote the blog post I wanted to write! (But you got some things wrong.)
> But it’s strictly worse than a regular bounty system.
Nope. In a normal bounty system, you set a bounty and if anyone is willing to do it for that price, they do the job and get the bounty. If they were willing to do the job for less than the bounty you set, they get the excess as profit.
Action markets have one advantage and one difference (which might be an advantage).
The advantage is that they allow decentralized claiming of the bounty:
Someone can buy up all the "yes" shares, clean half the backyard, then offer to sell them at the the halfway mark between the price they bought and 1, and thereby claim half the bounty. Moreover, in a sufficiently liquid market, the price that they get for their shares will reflect the market consensus on how much their action contributed to the goal (so you'd get less than half the bounty for cleaning the easier half).
The difference is that in normal bounties, the excess bounty (over what people are willing to do the job for) is paid to the person taking the bounty. In action markets, that excess is paid to people who are making actuate predictions and giving information to the market. I actually think this is pretty cool (but maybe you'd prefer to tip the people who perform your bounty rather than paying that money out to the people who are teaching you that you're wrong about how much it costs to get something to happen in the world?)
> For one thing, if I suspect someone else will take care of the backyard, then just by registering that prediction I can share the winnings with them, even though I didn’t do any work
You can, but only if you believe that sharing winnings with them will not push the bounty below what they'd be willing to take. And if you do that, you're providing information to the market.
For example, say there's 1000 shares of yes that can be bought for $0.70. This means there's a $300 bounty on cleaning the backyard. If you think someone is about to clean the backyard and try to preempt then by buying half the shares, you decrease the bounty to $150. If that's still enough for them to clean the backyard, that means the bounty-setters were miscalibrated on the bounty required to clean the backyard, and you're claiming the excess. If it's not enough, then you're betting incorrectly, and the price should not rise.
> (since this is at 70% (rather than 100%), someone must be doing something like this).
This is not true. If the probability is at 100% there's no bounty to be had. (Maybe you meant "rather than 0%"?). Additionally, *you can't decouple the action from the prediction*! You're supposed to set up an action market at a price $p and volume N shares such that you believe $(1-p)N is the minimum bounty required to make the action happen with probability p. If you differ from this constraint at all you will lose money (in a sufficiently liquid action market) to people who are better than you at predicting.
> For another thing, it incentivizes people to bet “no” and then sabotage cleanup projects.
Only if you set up your market wrong. If you're trying to make "yes" happen, you're not supposed to offer any shares of "no".
Importantly, none of the money that you are bountying towards "yes" can be extracted by the people who are trying to sabotage you. If the the yard gets cleaned, you pay out the bounty to the bounty-taker and the people who predicted better than you. If the yard doesn't get cleaned, you get compensated.
So insofar as there is an incentive to "bet no" and sabotage the yard cleaning, it must come from something other than your bounty (maybe it comes from the people playing the prediction game on your bounty market? maybe it comes from making the outcome and the current prediction more legible?)
> Banned because Wall Street developed a financial instrument that let them short ConTrackeds, then tried really hard to prevent bridges from being built
Ha, but this isn't an argument against action markets, it's an argument against inadequate equilibria. Why couldn't Wall Street come up with a financial instrument that lets them short bridge building directly and then go interfere in bridge building? If the shorting method involved literal borrowing of ConTrackeds, then this is an argument for the stupidity of the lenders, who should have realized that lending this token incentivizes the lendee to sabotage the project. At worst, you're not supposed to ban the action market, you're supposed to ban the derivative markets that incentivize antisocial behavior.
> Commenters brought up that you get weird incentives as soon as more than one person owns ConTrackeds
Link? The incentives, as I understand it, are:
1. If you bought the token to claim the bounty, you have an incentive to make partial progress on building the bridge, and to sell as soon as the market adjusts to account for the partial progress you believe you will make.
2. If you bought the token because you thought the market was wrong and you know better, you have an incentive to sell as soon as either the market adjusts or you learn otherwise.
3. You are incentivized to buy only if you either think you know better than the market or you are planning to make progress on the bridge building.
> I think you could solve this problem by subsidizing a prediction market at 25% chance of recovery, and letting doctors bet on other odds.
Yes, you can. This is what my formula above says in general.
> This would still have some problems - if someone saw a famous doctor betting, they could bet too, and “steal” some of the “winnings” from the doctor - but it seems to be kind of the right idea.
No, this means that either you were dumb and set a higher bounty than necessary (the doctor still buys enough shares to make it worth it to treat you) or that these people are dumb and will lose money to the market (there isn't enough bounty left to incentivize the doctor).
> Different prediction markets “leak” into being action markets at a different rate: one about when a distant star will supernova is 100% prediction; one about when a President will die is 99% prediction; one about when the forecasters’ backyard will be cleaned is maybe 80% action.
Important question: can you read off the leakage rate for a given market from market data? I think solving this problem is required for solving AI interpretability, based on the connection you mentioned.
Important follow-up: If you can't read it from market data but you can read it from brains / predictive processing / action-inference, what does this say about the design of brains, and can we port this design back over into the marketplace to make better prediction-action markets?
And finally, I want to note that this issue is already present in existing AI and prediction markets. The book Surveillance Capitalism talks about "behavioral futures markets" that Google and all the other big Internet companies run (I'm not sure how much this is a literal thing and how much it's a model), which are used for advertising. The book also talks about how they want to control our actions to make them predictable. (I think this book is secretly about misaligned super-human AI that is actually already here, and how we really better get our act together and align it before it's too late (unless it already is), except the author doesn't quite realize this. (I want to write a book review on it, or collaborate with someone on writing a book review on it.))
Final thoughts: there are five directions of incentives that action markets will incentivize:
1. They incentive better predictions. This is the standard incentive from prediction markets.
2. When there's much more money on selling "yes" (/ buying "no"), they incentivize "yes"
3. When there's much more money on selling "no" (/ buying "yes"), they incentivize "no".
4. When there is the same amount of money on both sides ("yes" and "no"), they incentivize whichever direction is easier to influence, i.e., they incentivize making behavior more predictable. This is the issue talked about in Surveillance Capitalism.
5. When the amount of money on each side, after being adjusted for the marginal cost of influencing the outcome in that direction, is the same, I *think* the market disincentivizes any action that shifts the likelihood away from the current prediction.
So beware that playing market maker and adding liquidity to prediction-action markets is not morally neutral. You are incentivizing whichever outcome is easier to cause. Since destruction is generally easier than creation, this says that we should tax market makers of action markets at the rate of destruction their market-making is liable to induce. (This is a great use case of being able to read the leakage rate off of market data! If we can read off the leakage rate, then we can accurately tax market makers!)
“The agreed-on solution for this is lots of conditional prediction markets.” - a downside here is that each individual market ends up being lower confidence, as bettors have to split up their time across many questions. But if you think that there are diminishing returns to time spent on a question, this is probably am overall win?
For the pregnancy market, you are assuming women are already taking prenatal vitamins, right? (Otherwise folic acid or iodine or something will be best).
> when I was a poor medical student, Zvi Mowshowitz’s company ran a contest for the best literature review of mineral supplementation with a $5000 prize. I put in 20 hours of research and won. If I’m typical, this is proof of concept that big enough bounties can incentivize poor medical students to do good lit reviews
Is there a website which aggregates links to these sorts of contests? (It could attract entrants who are indifferent to anything beyond the semblance of accuracy, so I wouldn't be surprised if biomedical companies wouldn't actually use it.)
Had you expected a large number of entrants, would you still have entered the contest? If so, you were probably motivated by more than just the money.
About the idea with organising market to decide which papers should be sent to e.g. Scott Aaronson: isn't it just a non-expert version of peer-review process?
The question about the dead drummer is not a question about past events it is a question about a future event -- i.e. the contents of the autopsy report. If the report is not made public, it may actually be about what the ME says about the report.
in 2017 someone bombed bus of top German soccer team, having “allegedly bought options to short sell 15,000 shares of Borussia Dortmund stock for 78,000 euros [£65,000]".
It seems like some of the moral hazard issues with prediction markets would be well addressed by limiting the players and the potential stake of any players. Even if we assume that someone might be willing to do antisocial things for money, the effort involved should at least make low-stakes bidding possible. ... until you get multiple, redundant markets, that is.
How about an attention market where the currency is attention of every participant?
Every person upvotes content that they think deserves your attention and downvotes content that will waste your attention. They have some starting balance of your attention tokens and with every upvote/downvote they place a bet. If you upvote some content - you reward everyone who upvoted it by giving them 100 tokens of your attention (divided equally). And you penalize those who downvoted it by taking away some percentage of your attention tokens they hold.
You do the opposite when you downvote content - penalize those who upvoted and reward those you downvoted.
Then in such system you can rank content based on the amount of your attention token bets and the content that is ranked at the top naturally gets your attention.
This creates a feedback loop where you give your attention to those who have earned your attention in the past.
I am building such a system at linklonk.com and would appreciate any feedback.
Note on crackpots: you do not, in fact, need to be Scott Aaronson. In my experience, any cosmology PhD student who's put anything on arXiv will receive a good amount of "refutations" of all of modern physics!
I was waiting impatiently for the active inference punchline, which I think might be rather profound, insofar as you've just illustrated that markets are virtual organisms.
Thanks for the writeup, Scott! I've honestly been super thrilled to see all the different use cases our community has figured out for the prediction market mechanism; I think it really validates the approach of letting people create markets on anything they want!
One more example of a bounty/action market that worked pretty well: Elizabeth & I posted a $100 bounty for some data aggregation work for Slime Mold Time Mold's hypothesis of contamination-led obesity. It got two separate takers: https://manifold.markets/Austin/100-bounty-will-someone-combine-the
Would love to figure out how to better support this natively -- if anyone has suggestions, please reach out!
(1) "What will I (Scott Alexander) rate as most promising when I do a deep dive into the research on pregnancy interventions?"
I'm going for the "Natural birth versus Caesarian"
(2) "They’re betting on whether someone will clean up the (currently disastrous) backyard. They’ve said pretty openly that they’re hoping someone will buy a lot of “yes”, take care of the backyard project, and then take all their money. "
Could they not just pool the money and hire a landscaper or groundskeeper? When my back garden was a disaster, after a long time dithering about "something needs to be done", we finally hired a guy to come cut the hedges, cut down and take away the tree that was growing in the wrong place, and do a ton of weeding and spraying. Result! If you're waiting for someone else to do it, it will never get done - "what's everybody's business is nobody's business".
Is the literature review of mineral supplementation public?
I'm not sure how the action market helps with the doctor situation. Isn't creating an action market for it just worse than the solution of asking a doctor to help, paying him if it does and demanding that he pay you if it does not? Your complaint with that system seemed to be that a risk-adverse doctor who was less than 100% confident might pass, but it seems like the action market would have the same issue here.
With respect to the Scott Aaronson thing, my understanding is that the American Math Association has come up with a solution. Or maybe some other math organization; it was explained to me a couple years ago, so one or two details may be off.
Most people who submit proofs for the Riemmann Hypothesis (or any other similarly famous problem) are are not so correct that a PhD student can't take apart their argument. So, the AMA charges about $100 per page of the proof, and then gives it to a PhD student to dissect. The student writes up a clear report on why the proof doesn't work, which is then returned to the author, and the student gets some cut of the money to use as travel funds.
Now, this may not still be totally accurate, or perhaps the person once telling me it was not accurate in their description, but it still feels like a good system.
Two markets this week:
An action market - if I keep dumping money on it, hopefully somebody can make it happen: https://manifold.markets/AlexPower/will-thestalwart-have-a-verified-tw
And a "no-action" market, more specifically a prediction market where it seems nobody knows the answer: https://manifold.markets/AlexPower/will-the-country-of-turkey-be-recog
> Niels Bohr supposedly said that “prediction is very difficult, especially about the future”.
I would have guessed that was a Yogi Berra quote:
• "When you come to a fork in the road, take it."
• On economics: "A nickel ain't worth a dime anymore."
• "Always go to other people's funerals. Otherwise, they won't come to yours."
• And last but certainly not least: "It's deja vu all over again."
>I once joked that instead of lower courts, we should have prediction markets on what the Supreme Court would think of a given case. Same idea, higher stakes.
Related idea I've had whenever there's a SCOTUS vacancy: a major qualification for appellate court judges to be elevated should be the extent to which they have exercised their delegated authority faithfully, which is to say, when cases they've decided subsequently make it before the highest court, do the Justices agree with them? 5-4 splits would have the least weight either way, 9-0 agreeing/disagreeing with the judge would have the greatest favorable/unfavorable weight; more emphasis given to cases in which the judge was in the dissent at the Circuit.
Isn't the main difference for incentivizing actions via the stock market vs. prediction markets is that stock markets are highly regulated and large transactions such as placing a major short on Tesla the day before Elon Musk is assassinated would be easily tracked by the SEC/law enforcement. The companies running prediction markets could presumably track these types of things and inform law enforcement about them, but would they necessarily have the infrastructure/processes in place to do so? And would they all voluntarily do that? Surely some privacy libertarian style prediction market would advertise how it doesn't track any user data or supply it to the govt. And what about blockchain prediction markets using anonymous untraceable cryptocurrencies like Monero?
Most other major markets where you could make large sums of money by taking a huge position and purposely affecting change in that market has some form of regulatory overview that presumably helps deter these kinds of things.
I know it was said as a joke but just to clarify, lower courts are not intended to have the same decisions as appellate courts.
Note that, while it's obviously not for money, fantasyscotus.net is quite real and could quite easily be monetized in a different regulatory environment.
For many of these applications, I think one of the biggest challenges for prediction markets will be to beat the crude approximations to prediction markets that we already self-organize on an ad-hoc basis. To take an example from this post where I have some firsthand knowledge, I indeed don't read most of the P=NP proofs, refutations of quantum computing, and other revolutionary developments that cross my field of attention -- but that can be overridden by colleagues I trust staking some of their reputation that, hey, this is interesting, this is worth looking at, etc. Which means that a prediction market wouldn't just have to beat nothing whatsoever, it would have to beat THAT.
In the meantime, though, if it's really important to optimize my use of time, then there's surely lower-hanging fruit to be had, like me not spending hours per day laying in bed with my phone scrolling through social media like now... :-)
The phrase "the market is wrong" makes no sense to me.
Consider a market created by punters on a horse race. If the longshot going off at 40-1 odds wins, it does not mean the market was wrong. That horse *should* win about 1/40 races. OTOH, if the longshot wins 40 races in a row, that market probably isn't functioning. More important to note, though, is that if the favorite wins 40 races in a row, that also means the market isn't functioning.
If you want to know how well a market is functioning, give the market a Brier score across at least dozens of predictions. The EMH is true by tautology if you have a true market.
This sentence has a minor error where the order of 2 words is reversed, but it happens to be an error that tricked my internal parser into going down a garden path that actually took me a nontrivial amount of time to back out of:
"I throw these out all unread the time when I get them in the mail"
I considered doing the fix Valinor's backyard myself thing, but (a) it was unclear to me that they were actually angling for that and I thought it might be intrusive if I just showed up and started touching their stuff, and (b) if I'm going to do manual labor I want to be paid in real actual money and not Monopoly money ¯\_(ツ)_/¯
Cytomegalovirus should definitely feature in the biodeterminist's guide to parenting. I don't see it anywhere in that list.
On why we don't have more geniuses: "His summer following his father around learning medicine was probably good for him, but not outside the bounds of what still happens today (I followed my father around learning medicine)."
On whether small prizes are great for thoughtful, complex: "I put in 20 hours of research and won. If I’m typical, this is proof of concept that [...]"
This is maybe not the most effective argument that parental tutoring is less than very effective, or that medical students can be cheaply incentivized to do good lit reviews.
Don't wait too long to do your "deep dive" on pregnancy interventions, so that my wife and I can benefit from it in the (somewhat) near future! :)
> People don’t assassinate Elon Musk because then they’d be investigated for murder.
Well that's an interesting quote.
In regard to the historical prediction market, I wonder how you could use them to judge the consensus on past events. There are many things that happened in the past that are still unclear, with papers being published explaining that this and this happened and two months later a counter-paper that says it totally did not. It is often hard for non-experts to
1/ judge what is the most plausible explanation of what happened,
2/ judge how strong the consensus is.
I am starting this Manifold market with the first disputed event that came to my mind: did Joseph Stalin died of natural cause? (https://manifold.markets/TaoSumer/did-stalin-die-of-natural-cause)
It will be probably be too obscure, too niche, too complex and too irrelevant for it to gain traction (let me know if you got better ideas for a disputed event, which should be easy to find). I thought of using the same technique in my previous experimental market, as in "Convince me that XXX is true and I'll resolve the market this way" (https://manifold.markets/TaoSumer/are-there-more-doors-than-wheels-in), but I feel like such a world-changing event is too complex to ask people to debate constructively in the comments. I guess the best way to make it interesting would be to invite historians who wrote papers on Stalin deaths to bet money and debate/argue, and then get the crowd to decide who got the best arguments - tho rethoric would probably come in the way, with people voting for the best speaker/writer.
Do you have any ideas on how we can make such markets better? I can implement them for the Stalin market - or create a more interesting one.
>> Will Valinor unfuck its backyard eating area by April 1st.
Man I thought Jeopardy rules applied. Or it's a trick on two levels.
Oh man, you wrote the blog post I wanted to write! (But you got some things wrong.)
> But it’s strictly worse than a regular bounty system.
Nope. In a normal bounty system, you set a bounty and if anyone is willing to do it for that price, they do the job and get the bounty. If they were willing to do the job for less than the bounty you set, they get the excess as profit.
Action markets have one advantage and one difference (which might be an advantage).
The advantage is that they allow decentralized claiming of the bounty:
Someone can buy up all the "yes" shares, clean half the backyard, then offer to sell them at the the halfway mark between the price they bought and 1, and thereby claim half the bounty. Moreover, in a sufficiently liquid market, the price that they get for their shares will reflect the market consensus on how much their action contributed to the goal (so you'd get less than half the bounty for cleaning the easier half).
The difference is that in normal bounties, the excess bounty (over what people are willing to do the job for) is paid to the person taking the bounty. In action markets, that excess is paid to people who are making actuate predictions and giving information to the market. I actually think this is pretty cool (but maybe you'd prefer to tip the people who perform your bounty rather than paying that money out to the people who are teaching you that you're wrong about how much it costs to get something to happen in the world?)
> For one thing, if I suspect someone else will take care of the backyard, then just by registering that prediction I can share the winnings with them, even though I didn’t do any work
You can, but only if you believe that sharing winnings with them will not push the bounty below what they'd be willing to take. And if you do that, you're providing information to the market.
For example, say there's 1000 shares of yes that can be bought for $0.70. This means there's a $300 bounty on cleaning the backyard. If you think someone is about to clean the backyard and try to preempt then by buying half the shares, you decrease the bounty to $150. If that's still enough for them to clean the backyard, that means the bounty-setters were miscalibrated on the bounty required to clean the backyard, and you're claiming the excess. If it's not enough, then you're betting incorrectly, and the price should not rise.
> (since this is at 70% (rather than 100%), someone must be doing something like this).
This is not true. If the probability is at 100% there's no bounty to be had. (Maybe you meant "rather than 0%"?). Additionally, *you can't decouple the action from the prediction*! You're supposed to set up an action market at a price $p and volume N shares such that you believe $(1-p)N is the minimum bounty required to make the action happen with probability p. If you differ from this constraint at all you will lose money (in a sufficiently liquid action market) to people who are better than you at predicting.
> For another thing, it incentivizes people to bet “no” and then sabotage cleanup projects.
Only if you set up your market wrong. If you're trying to make "yes" happen, you're not supposed to offer any shares of "no".
Importantly, none of the money that you are bountying towards "yes" can be extracted by the people who are trying to sabotage you. If the the yard gets cleaned, you pay out the bounty to the bounty-taker and the people who predicted better than you. If the yard doesn't get cleaned, you get compensated.
So insofar as there is an incentive to "bet no" and sabotage the yard cleaning, it must come from something other than your bounty (maybe it comes from the people playing the prediction game on your bounty market? maybe it comes from making the outcome and the current prediction more legible?)
> Banned because Wall Street developed a financial instrument that let them short ConTrackeds, then tried really hard to prevent bridges from being built
Ha, but this isn't an argument against action markets, it's an argument against inadequate equilibria. Why couldn't Wall Street come up with a financial instrument that lets them short bridge building directly and then go interfere in bridge building? If the shorting method involved literal borrowing of ConTrackeds, then this is an argument for the stupidity of the lenders, who should have realized that lending this token incentivizes the lendee to sabotage the project. At worst, you're not supposed to ban the action market, you're supposed to ban the derivative markets that incentivize antisocial behavior.
> Commenters brought up that you get weird incentives as soon as more than one person owns ConTrackeds
Link? The incentives, as I understand it, are:
1. If you bought the token to claim the bounty, you have an incentive to make partial progress on building the bridge, and to sell as soon as the market adjusts to account for the partial progress you believe you will make.
2. If you bought the token because you thought the market was wrong and you know better, you have an incentive to sell as soon as either the market adjusts or you learn otherwise.
3. You are incentivized to buy only if you either think you know better than the market or you are planning to make progress on the bridge building.
> I think you could solve this problem by subsidizing a prediction market at 25% chance of recovery, and letting doctors bet on other odds.
Yes, you can. This is what my formula above says in general.
> This would still have some problems - if someone saw a famous doctor betting, they could bet too, and “steal” some of the “winnings” from the doctor - but it seems to be kind of the right idea.
No, this means that either you were dumb and set a higher bounty than necessary (the doctor still buys enough shares to make it worth it to treat you) or that these people are dumb and will lose money to the market (there isn't enough bounty left to incentivize the doctor).
> Different prediction markets “leak” into being action markets at a different rate: one about when a distant star will supernova is 100% prediction; one about when a President will die is 99% prediction; one about when the forecasters’ backyard will be cleaned is maybe 80% action.
Important question: can you read off the leakage rate for a given market from market data? I think solving this problem is required for solving AI interpretability, based on the connection you mentioned.
Important follow-up: If you can't read it from market data but you can read it from brains / predictive processing / action-inference, what does this say about the design of brains, and can we port this design back over into the marketplace to make better prediction-action markets?
And finally, I want to note that this issue is already present in existing AI and prediction markets. The book Surveillance Capitalism talks about "behavioral futures markets" that Google and all the other big Internet companies run (I'm not sure how much this is a literal thing and how much it's a model), which are used for advertising. The book also talks about how they want to control our actions to make them predictable. (I think this book is secretly about misaligned super-human AI that is actually already here, and how we really better get our act together and align it before it's too late (unless it already is), except the author doesn't quite realize this. (I want to write a book review on it, or collaborate with someone on writing a book review on it.))
Final thoughts: there are five directions of incentives that action markets will incentivize:
1. They incentive better predictions. This is the standard incentive from prediction markets.
2. When there's much more money on selling "yes" (/ buying "no"), they incentivize "yes"
3. When there's much more money on selling "no" (/ buying "yes"), they incentivize "no".
4. When there is the same amount of money on both sides ("yes" and "no"), they incentivize whichever direction is easier to influence, i.e., they incentivize making behavior more predictable. This is the issue talked about in Surveillance Capitalism.
5. When the amount of money on each side, after being adjusted for the marginal cost of influencing the outcome in that direction, is the same, I *think* the market disincentivizes any action that shifts the likelihood away from the current prediction.
So beware that playing market maker and adding liquidity to prediction-action markets is not morally neutral. You are incentivizing whichever outcome is easier to cause. Since destruction is generally easier than creation, this says that we should tax market makers of action markets at the rate of destruction their market-making is liable to induce. (This is a great use case of being able to read the leakage rate off of market data! If we can read off the leakage rate, then we can accurately tax market makers!)
> Or by going long on train companies, and bombing a plane?
I assume you mean train, not plane here.
“The agreed-on solution for this is lots of conditional prediction markets.” - a downside here is that each individual market ends up being lower confidence, as bettors have to split up their time across many questions. But if you think that there are diminishing returns to time spent on a question, this is probably am overall win?
For the pregnancy market, you are assuming women are already taking prenatal vitamins, right? (Otherwise folic acid or iodine or something will be best).
> I don’t think there’s any site as convenient as Mantic
Should be Manifold I think
> when I was a poor medical student, Zvi Mowshowitz’s company ran a contest for the best literature review of mineral supplementation with a $5000 prize. I put in 20 hours of research and won. If I’m typical, this is proof of concept that big enough bounties can incentivize poor medical students to do good lit reviews
Is there a website which aggregates links to these sorts of contests? (It could attract entrants who are indifferent to anything beyond the semblance of accuracy, so I wouldn't be surprised if biomedical companies wouldn't actually use it.)
Had you expected a large number of entrants, would you still have entered the contest? If so, you were probably motivated by more than just the money.
About the idea with organising market to decide which papers should be sent to e.g. Scott Aaronson: isn't it just a non-expert version of peer-review process?
One quibble: do you think the comment you made about the pregnancy interventions *won't* interfere with the betting?
The question about the dead drummer is not a question about past events it is a question about a future event -- i.e. the contents of the autopsy report. If the report is not made public, it may actually be about what the ME says about the report.
in 2017 someone bombed bus of top German soccer team, having “allegedly bought options to short sell 15,000 shares of Borussia Dortmund stock for 78,000 euros [£65,000]".
https://www.theweek.co.uk/83549/borussia-dortmund-attack-stock-market-trader-arrested
It seems like some of the moral hazard issues with prediction markets would be well addressed by limiting the players and the potential stake of any players. Even if we assume that someone might be willing to do antisocial things for money, the effort involved should at least make low-stakes bidding possible. ... until you get multiple, redundant markets, that is.
Can market fidelity be a thing?
The 'starting a prediction market for whether the yard will be cleaned up' strikes me as the (far less dysfunctional) rationalist analogue of this: https://www.theonion.com/marxists-apartment-a-microcosm-of-why-marxism-doesnt-wo-1819566655
How about an attention market where the currency is attention of every participant?
Every person upvotes content that they think deserves your attention and downvotes content that will waste your attention. They have some starting balance of your attention tokens and with every upvote/downvote they place a bet. If you upvote some content - you reward everyone who upvoted it by giving them 100 tokens of your attention (divided equally). And you penalize those who downvoted it by taking away some percentage of your attention tokens they hold.
You do the opposite when you downvote content - penalize those who upvoted and reward those you downvoted.
Then in such system you can rank content based on the amount of your attention token bets and the content that is ranked at the top naturally gets your attention.
This creates a feedback loop where you give your attention to those who have earned your attention in the past.
I am building such a system at linklonk.com and would appreciate any feedback.
Note on crackpots: you do not, in fact, need to be Scott Aaronson. In my experience, any cosmology PhD student who's put anything on arXiv will receive a good amount of "refutations" of all of modern physics!