In previous discussions of optimal investing, the efficient market hypothesis has been repeatedly cited to say that you cannot easily predict the market, and that even if you could, you would be better off working for a bank, rather than risking your own money.
But suppose I had discovered an entirely novel technique for predicting the market, one sufficiently complex that the EMH does not apply. How would I get a job at a bank to leverage this?
I could show them my technique, but then (in the unlikely event they took me seriously) they could copy it. I could get them to sign a non-disclosure agreement, but what mechanism is there to enforce such an agreement, when the technique could be used behind closed doors?
I could track my predictions, registered with some impartial third party, but then if thousands of people did this (which they do) or one person did this with thousands of sockpuppets, then some of these people would show extremely good results by sheer luck.
One reason this may be important is that ‘superhuman finance AI’ is strictly simpler than ‘superhuman AGI’ and it’s important to work out how easy it is for AI to acquire massive resources by dominating the world’s financial markets.
META: I considered posting this to the comments on the last optimal investing post, but since its several months old I assumed few would read it.
Well, you could invest your own money. Most strategies benefit from using small amount of cash, since being able to take advantage of small volume opportunities often trumps higher relative execution costs. Register your track record and start recruiting investors.
You can also trade with your own strategies at companies that let you use a mis of your money and theirs, and provide some resources. Typically, traders at these companies aren’t executing pure softwares strategies, so “stealing” their methods doesn’t have much point.
You could also go to people with established reputations in academic or quantitative finance and show them enough of your method to convince them you’re legit, but not so much they can copy it, then have them lend their reputation.
A superhuman AGI could accumulate tremendous financial resources; to do so most effectively it would need access to as many feeds and exchanges as possible, so some kind of shell company would buy those for it. I’m not sure to what extent a finance-specialist AI that achieves superhuman performance is really easier to make than a superhuman AGI; and bear in mind that it could lose its edge quickly.
You can also trade with your own strategies at companies that let you use a mis of your money and theirs, and provide some resources.
I’m not sure I follow you, do you mean leverage or something else?
You could also go to people with established reputations in academic or quantitative finance and show them enough of your method to convince them you’re legit, but not so much they can copy it, then have them lend their reputation.
Any information about a method narrows the hypothesis space, so I’m not sure this is possible. But discussion with academics, who presumably are not actually trading is an interesting idea.
I’m not sure to what extent a finance-specialist AI that achieves superhuman performance is really easier to make than a superhuman AGI
Its not at all obvious whether its easier or not—if AI can be broken down into subsystems, such as “planning” “hypothosis generation” and “predictions” then it is a matter of solving just the prediction subsystem, which is of course easier than solving all the other problems as well. OTOH, “hypothosis generation” could be an integral part of prediction, and the subsystems might not be able to operate independently.
even if you could, you would be better off working for a bank, rather than risking your own money.
That depends on the particulars. In any case, this is more the bailiwick of hedge funds rather than banks nowadays.
suppose I had discovered an entirely novel technique for predicting the market
You would have to demonstrate that it works and convince people with money to give some of that money to you to trade.
You can try to persuade people with just your technique (and backtesting results) under an NDA, but without a track record it will be hard—or the deals offered to you won’t be good. You also can trade your own money—in whatever amounts you have—for a while in a separate audited account. A good track record and a demonstrated willingness to risk your own money on your idea will help you persuade other people that the idea works.
Remember, you don’t have to mathematically prove anything, all you have to do is convince people with money to give you some :-)
how easy it is for AI to acquire massive resources by dominating the world’s financial markets.
For a fully developed AI it will be easy, but I am not sure why would it bother. And if you’re thinking about trading “AIs” developed on Wall St. and such, these are very likely to be narrow tool-like AIs with very specific and limited goals.
And if you’re thinking about trading “AIs” developed on Wall St. and such, these are very likely to be narrow tool-like AIs with very specific and limited goals.
Alexander Wissner-Gross disagrees with you. He believes “superhuman AI” will emerge from quantitative finance. His talk at Singularity Summit is below.
In previous discussions of optimal investing, the efficient market hypothesis has been repeatedly cited to say that you cannot easily predict the market, and that even if you could, you would be better off working for a bank, rather than risking your own money.
But suppose I had discovered an entirely novel technique for predicting the market, one sufficiently complex that the EMH does not apply. How would I get a job at a bank to leverage this?
I could show them my technique, but then (in the unlikely event they took me seriously) they could copy it. I could get them to sign a non-disclosure agreement, but what mechanism is there to enforce such an agreement, when the technique could be used behind closed doors?
I could track my predictions, registered with some impartial third party, but then if thousands of people did this (which they do) or one person did this with thousands of sockpuppets, then some of these people would show extremely good results by sheer luck.
One reason this may be important is that ‘superhuman finance AI’ is strictly simpler than ‘superhuman AGI’ and it’s important to work out how easy it is for AI to acquire massive resources by dominating the world’s financial markets.
META: I considered posting this to the comments on the last optimal investing post, but since its several months old I assumed few would read it.
Well, you could invest your own money. Most strategies benefit from using small amount of cash, since being able to take advantage of small volume opportunities often trumps higher relative execution costs. Register your track record and start recruiting investors.
You can also trade with your own strategies at companies that let you use a mis of your money and theirs, and provide some resources. Typically, traders at these companies aren’t executing pure softwares strategies, so “stealing” their methods doesn’t have much point.
You could also go to people with established reputations in academic or quantitative finance and show them enough of your method to convince them you’re legit, but not so much they can copy it, then have them lend their reputation.
A superhuman AGI could accumulate tremendous financial resources; to do so most effectively it would need access to as many feeds and exchanges as possible, so some kind of shell company would buy those for it. I’m not sure to what extent a finance-specialist AI that achieves superhuman performance is really easier to make than a superhuman AGI; and bear in mind that it could lose its edge quickly.
I’m not sure I follow you, do you mean leverage or something else?
Any information about a method narrows the hypothesis space, so I’m not sure this is possible. But discussion with academics, who presumably are not actually trading is an interesting idea.
Its not at all obvious whether its easier or not—if AI can be broken down into subsystems, such as “planning” “hypothosis generation” and “predictions” then it is a matter of solving just the prediction subsystem, which is of course easier than solving all the other problems as well. OTOH, “hypothosis generation” could be an integral part of prediction, and the subsystems might not be able to operate independently.
That depends on the particulars. In any case, this is more the bailiwick of hedge funds rather than banks nowadays.
You would have to demonstrate that it works and convince people with money to give some of that money to you to trade.
You can try to persuade people with just your technique (and backtesting results) under an NDA, but without a track record it will be hard—or the deals offered to you won’t be good. You also can trade your own money—in whatever amounts you have—for a while in a separate audited account. A good track record and a demonstrated willingness to risk your own money on your idea will help you persuade other people that the idea works.
Remember, you don’t have to mathematically prove anything, all you have to do is convince people with money to give you some :-)
For a fully developed AI it will be easy, but I am not sure why would it bother. And if you’re thinking about trading “AIs” developed on Wall St. and such, these are very likely to be narrow tool-like AIs with very specific and limited goals.
Alexander Wissner-Gross disagrees with you. He believes “superhuman AI” will emerge from quantitative finance. His talk at Singularity Summit is below.
https://www.youtube.com/watch?v=p4amJyOfbec