AI & Trading

Hello All,

A friend of mine from college recently opened a trading company using a kind of artificial intelligence that he developed. It supposedly inductively learns rules from a given set of data. I saw the outputs and they are very very impressive. However, before I commit money, I want to ask you knowledgeable people whether any of you has ever used AI in trading? If so, how did you find it?

 

be careful.
although I am very pro automated trading and use it myself, AI systems have a major flaw. They use past data to extrapolate the future. Although some argue trading is about that, I disagree.
Trading is about placing bets with good odds while following the market or going against it.
AI systems (neural nets since that's what your friend seems to use) are overoptimized systems on past data. They usually do not do well in actual trading.

Maybe, you never know, your friend found a way to train his machine in a way it learned about basic market mechanics (which obviously exists). In that case it is a little better, but I doubt it.

One thing you MUST do is try the system in forward testing LIVE. so with real time data (or end of day, if you enter at open or close only). in other words, you must papertrade enough to see if the system works in real life as well as in back test.
well, and if the system already seems to work in LIVE mode too, there are many other things to be aware of... A system is as good as its worst bad streak. So you must know how bad historically the system did (drawdown, account volatility etc..) and what is the risk of ruin for the system (what does it take to wipe you out).
Without these, so called good results mean nothing.

That's a difficult thing with AI systems, they are often difficult to backtest because software is rarely made for that.
Well, the list of points would be long.. let's say I just gave you some pointers, you need much more data before having good reasons to commit money.

tntneo

 

Thanks for the suggestions. But here's what I do know.

He trained his machine (not a neural net) with historical data. Then he used the results to backtest on data that weren't included in the training data. So the machine never learned those. The result is that the machine is accurate about the directional changes in the S&P one month forward at 90% of the time. Unless he's fooling me by using a trick program (which I highly doubt), results like that can't be ignored. What do you think?

 

AI IS THE FUTURE! Smart machines will soon out perform their human creators. Please post your friend's website if he has one. Give us more info...very fascinating. machine will soon out think and can react infinitely faster and over a larger scale than their human counterparts. Let's hear more...

 

Well, I can't really tell if your post was sarcasm or not. Maybe it's just my paranoia.

Anyway, I'm bounded by confidentiality not to give specific information about the AI. As you can imagine, he doesn't want something potentially this power to get out.

But I can definately say that it's nothing SO revolutionary. The pieces of the technology can be found in the AI systems that large trading firms have used over the years. He just found a way to cleverly combined various aspects of existing technologies to make something a bit more powerful.

The problem is, however, coming out with results that work but don't make a whole lot of logical sense. Is it better to trust machine learning, throw away human intuition in favor of AI intuition, or simply scrap the AI?

 

tnt,

Your argument is flawed. Every human trader uses historical data to extrapolate the future. However, the rub is machines will soon crunch those trading algorithms much faster and with more depth and on a much larger scale than you could ever hope for. Your assertion that history plays none or very little role in your trading is total garbage! History is all you have! It is the sum total of your market experience and knowledge...you are not creating trading scenarios/strategies out of thin air!!arrrrrrrrrrrrggggggghh!

 

JS11374: “The result is that the machine is accurate about the directional changes in the S&P one month forward at 90% of the time.”

Being correct about directional changes doesn’t tell you much about its long-term profitability. The system's P/L and winning percentage are completely dependent on it’s trade management parameters. How does the system manage its positions?

1) What constitutes being “accurate” about a directional change? A tick or more?

2) At what point does the system deem a trade “inaccurate” or how far can the position travel against you and still be labeled “accurate”?

• Does it achieve its percentage of accuracy by allowing a prohibitive amount of risk (i.e. large stop)?
  
• What type of drawdown will you experience the other 10% of the time?[/list=a]You may already be aware of the answers to these questions and find them to have been satisfactorily answered. No need to respond to my questions. Just food for thought. Good luck.

 

See, you are right. The problem (not a flaw, but just a limitation of the AI) is that the AI doesn't tell you precisely when to enter a position. In fact, it only tells you whether the S&P will rise, remain neutral, or fall next month.

That said, the error (times and magnitude that the machine predicted an upturn but got a down turn, and vice versa) evens out. Of course it doesn't address exactly how to trade with that.

Speaking of which, if you know which direction the market will go next month, but dont' know by how much or when, how would you trade with it?

 

we have used neural nets and genetic algorithms at various stages of our research, yet wound up trading a simple top-down system that has very little to do with AI. the 'nets were just too unreliable and too prone to overoptimizing.

anyhow, no matter how good your friend's system is, you MUST be prepared to lose all the money you give him.

- jaan

 

JS,

go here for more insight.

http://clubs.yahoo.com/clubs/investingwithdeepinsight

i believe you'll find the discussion very interesting.

ps ignore the last few posts there...offtopic