Seems like the definition of “severe” is an issue here. Maybe I should have used “incredibly severe”?
Yes, reinsurance markets deal in large insured risks, but they do not target the incredibly large humanitarian risks that are more informative to us. See reinsurance deals here for reference: http://www.artemis.bm/deal_directory/
I don’t think they (or your proposed prediction markets) would be helpful in estimating the probabilities of extinction events.
Care to explain your reasoning? For example, if the market indicated that the chance of a pandemic killing 50% of the population is 1,000x greater than the likelihood of a nuclear war of any kind, wouldn’t a forecaster find this at least a little useful?
For the prediction markets to work they need to settle: a bet must be decided one way or another within reasonable time so that the winners could collect the money from the losers.
How are you going to settle the bets on a 50%-population pandemic or a nuclear war?
a bet must be decided one way or another within reasonable time
Each contract would have a maturity date—that is standard.
How are you going to settle the bets on a 50%-population pandemic or a nuclear war?
Your primary concern is that the market would not be functional after a 50%-population pandemic or a nuclear war? That is a possibility. The likelihood depends on the severity of the catastrophe, popularity of the market, its technology and infrastructure, geographic distribution, disaster recovery plan, etc.
With the proper funding and interest, I think a very robust market could be created. And if it works, the information it provides will be very valuable (in my opinion).
Each contract would have a maturity date—that is standard.
So, a bet would look like “There will or will not be a nuclear war during the year 2016”? I am not sure you will find enough serious bidders on the “will be” side to actually provide good predictions. You are likely to get some jokers and crackpots, but for prediction purposes you actually don’t want them.
Is there any empirical data that prediction markets can correctly estimate the chances of very-low-probability events?
...With the proper funding and interest, I think a very robust market could be created. And if it works, the information it provides will be very valuable (in my opinion).
...Is there any empirical data that prediction markets can correctly estimate the chances of very-low-probability events?
Liquidity problems are an issue but it may have been partially solved by first, paying normal interest on deposits to avoid opportunity cost issues, and second by market makers like Hanson’s LMSR. In particular, people can subsidize the market-maker, paying to get trading activity and hence accuracy.
It’s not the liquidity problems I’m worried about, but rather the signal-to-noise ratio.
Assume that the correct underlying probability is, say, 0.5% and so you should expect 0.5% of the participants to bet on the “the end is nigh!” side. However you also have a noise level—say 3% of your participants are looking for teh lulz or believe that a correct bet will get them front-row Rapture seats (or vice versa). Given this noise floor you will be unable to extract the signal from the prediction market if the event has a sufficiently low probability.
Assume that the correct underlying probability is, say, 0.5% and so you should expect 0.5% of the participants to bet on the “the end is nigh!” side.
? That’s not a prediction market. A PM is a continuous price interpretable as a probability, which traders trade based on its divergence from their estimated probability. You can buy or sell either way based on how it’s diverged. I have ‘bet for’ and ‘bet against’ things I thought were high or low probability all the time on past prediction markets.
You’re right, I got confused with the noise floor in polls.
However my concern with counterparties didn’t disappear. People need incentives to participate in markets. Those who bet that there won’t be a nuclear war next year will expect to win some money and if that money is half of a basis point they are not going to bother. Who will be the source of money that pays the winners?
At which probability you’re going to take the side “yes, there will be a nuclear war in 2016”?
And if this particular market will be subsidized, well, then it becomes free money and your prediction ability goes out of the window.
I suspect prediction markets won’t work well with very-low-probability bets. Especially bets with moral overtones (“You bet on a global pandemic happening? What are you, a monster??”)
Those who bet that there won’t be a nuclear war next year will expect to win some money and if that money is half of a basis point they are not going to bother.
Half a basis point is half a basis point. Bond traders would prostitute their grandmothers for a regular half a basis point boost to their returns.
Who will be the source of money that pays the winners?
The source is everyone who takes the other side of the contract? PMs are two-sided.
And if this particular market will be subsidized, well, then it becomes free money and your prediction ability goes out of the window.
I don’t pretend to understand how the LMSR and other market-makers really work, but I think that you can’t simply trade at random to extract money from them.
Especially bets with moral overtones (“You bet on a global pandemic happening? What are you, a monster??”)
Seems to have worked so far with the IARPA bets. (Admittedly, my own trades on ISIS being overblown didn’t work out too well but I think I did gain on the various flu contracts.)
Seems like the definition of “severe” is an issue here. Maybe I should have used “incredibly severe”?
Yes, reinsurance markets deal in large insured risks, but they do not target the incredibly large humanitarian risks that are more informative to us. See reinsurance deals here for reference: http://www.artemis.bm/deal_directory/
Care to explain your reasoning? For example, if the market indicated that the chance of a pandemic killing 50% of the population is 1,000x greater than the likelihood of a nuclear war of any kind, wouldn’t a forecaster find this at least a little useful?
For the prediction markets to work they need to settle: a bet must be decided one way or another within reasonable time so that the winners could collect the money from the losers.
How are you going to settle the bets on a 50%-population pandemic or a nuclear war?
Each contract would have a maturity date—that is standard.
Your primary concern is that the market would not be functional after a 50%-population pandemic or a nuclear war? That is a possibility. The likelihood depends on the severity of the catastrophe, popularity of the market, its technology and infrastructure, geographic distribution, disaster recovery plan, etc.
With the proper funding and interest, I think a very robust market could be created. And if it works, the information it provides will be very valuable (in my opinion).
So, a bet would look like “There will or will not be a nuclear war during the year 2016”? I am not sure you will find enough serious bidders on the “will be” side to actually provide good predictions. You are likely to get some jokers and crackpots, but for prediction purposes you actually don’t want them.
Is there any empirical data that prediction markets can correctly estimate the chances of very-low-probability events?
Liquidity problems are an issue but it may have been partially solved by first, paying normal interest on deposits to avoid opportunity cost issues, and second by market makers like Hanson’s LMSR. In particular, people can subsidize the market-maker, paying to get trading activity and hence accuracy.
It’s not the liquidity problems I’m worried about, but rather the signal-to-noise ratio.
Assume that the correct underlying probability is, say, 0.5% and so you should expect 0.5% of the participants to bet on the “the end is nigh!” side. However you also have a noise level—say 3% of your participants are looking for teh lulz or believe that a correct bet will get them front-row Rapture seats (or vice versa). Given this noise floor you will be unable to extract the signal from the prediction market if the event has a sufficiently low probability.
? That’s not a prediction market. A PM is a continuous price interpretable as a probability, which traders trade based on its divergence from their estimated probability. You can buy or sell either way based on how it’s diverged. I have ‘bet for’ and ‘bet against’ things I thought were high or low probability all the time on past prediction markets.
You’re right, I got confused with the noise floor in polls.
However my concern with counterparties didn’t disappear. People need incentives to participate in markets. Those who bet that there won’t be a nuclear war next year will expect to win some money and if that money is half of a basis point they are not going to bother. Who will be the source of money that pays the winners?
At which probability you’re going to take the side “yes, there will be a nuclear war in 2016”?
And if this particular market will be subsidized, well, then it becomes free money and your prediction ability goes out of the window.
I suspect prediction markets won’t work well with very-low-probability bets. Especially bets with moral overtones (“You bet on a global pandemic happening? What are you, a monster??”)
Half a basis point is half a basis point. Bond traders would prostitute their grandmothers for a regular half a basis point boost to their returns.
The source is everyone who takes the other side of the contract? PMs are two-sided.
I don’t pretend to understand how the LMSR and other market-makers really work, but I think that you can’t simply trade at random to extract money from them.
Seems to have worked so far with the IARPA bets. (Admittedly, my own trades on ISIS being overblown didn’t work out too well but I think I did gain on the various flu contracts.)