As a college undergrad (yes, I profess my status as a CS underling) doing a senior thesis in GA, I have read some papers on this GA's applied to trading. The dirty secret that these papers do not tell you is that: genetic algorithms are quite stochastic, meaning that they can evolve quite varying quality of technical trading rules, depending on the mood of the CS God during any particular evolution trials.
And these papers' results, in the interest of the authors who wrote them, will only tell you the best trial run that they have gotten - but the trouble is: out in the real world, judging your trading agent's performance is not by running hundreds of GA trials on historical stock data and publishing the best results; but you have gotta put your money where your GA technical rules tell you, whether the rules evolved are sub-optimal or not.
I spent a few months trying to go from "wow!" numbers in papers to something that I thought could be used for risking my own hard-earned money, and I failed. I think it's tilting at windmills unless you have the resources of Goldman Sachs or someone like that behind you.
I know for a fact that some investment banks use these techniques; they just aren't using them to make a mechanical "trade/no-trade" decision.
What Joe Schmoe on the street reading these papers is interested in is the possibility of lucrative personal trading. The papers ostensibly claim to instead be examining market efficiency. Consequently reading these papers is kind of disappointing. But, the reason I submitted the paper was its unorthodox approach which differs from most of the other literature on applying evolutionary computation to mechanical trading.
What you mean as "failed"? Even if you're bringing in stable (in sense of Sharpe ratio) returns above the standard interest rates, you've already succeed.
I don't think it makes sense to play with your money unless you're very rich. Returns like 50% per year are quite unlikely in the longer run, yet very risky. Instead, consider starting a fund.
The profit-loss patterns weren't good enough to be ready for real-life trading.
I'd be much more willing to tinker with it if I could do it with someone else's money. The hedge fund motto is "other people's lives, other people's money."
Correct. There can be permanent changes in market structure that make historical data irrelevant. The term for this is regime change. That and model risk.
Interesting, it is reasonable to suspect that many brokerage houses (with billions invested in R&D and very intelligent individuals) are using the same approaches to the problems with the added benefit of a human driver behind the wheel. It seems to me that purely using automated approaches to investing (in this case, GP market timing) will always be inferior to the insights of intelligent investors coupled with intelligent computing, but that alone would be a major shift from most investors current practices.
I am fascinated no end that a group of nerds is able to game the system to such an astounding degree. In this century, it isn't by gaming the casino that loads of cash can be made.
These guys run a hedge fund that is so successful it no longer accepts money from the public, and only manages the assets of its own employees. Using different secret strategies and to an extent human-supervised automatic trading. Its annualized return since its birth in the eighties has been on the order of 37%.
There are certainly plenty of firms that use AI-based investing techniques, but my understanding is that after being fashionable in the late 90s, they've fallen out of favor. After a couple AI-based systems "blew up," people realized that they're essentially opaque: you can't look at a neural network or a genetic algorithm and realize precisely what went wrong and fix it for next time. There's nowhere to place the blame -- not even on the programmer, since he didn't directly create the trading rules.
And these papers' results, in the interest of the authors who wrote them, will only tell you the best trial run that they have gotten - but the trouble is: out in the real world, judging your trading agent's performance is not by running hundreds of GA trials on historical stock data and publishing the best results; but you have gotta put your money where your GA technical rules tell you, whether the rules evolved are sub-optimal or not.