I was reading a blog entry on CPPTrader that was referencing an article from Bloomberg. I am a newbie with respect to trading, but with respect to AI I have been a few times around the block. It was something that I studied in University and have had an attraction to for a long time. I guess I am lazy and would love to write a program that “thinks” for me. As a sidenote I am using AI in my algorithmic trading software, but in a different context.
People should read Isacc Asimov more often. You might know about Isacc Asimov from the movie I,Robot featuring Wil Smith. In the movie Wil Smith plays a police detective that has an axe to grind with robots. The root of the problem is that a robot saved his life, and not a child’s because the probability of Wil’s character surviving were higher than that of the child. Wil’s character says that the robot misjudged because not everything can be expressed as a probability. And that point hits the core reason why AI as discussed in the article cannot work.
Recently, Jon Stewart interviewed Lee Gutkind on the Daily Show, and said the following.
Robots are so new so special so that if you take this ground hog robot… It’s also an incredibly frustrating experience because once you get a robot to do one thing well one time it does not necessarily mean that, that robot will do it again as beautifully.
I would like to sum it up and make the following statement.
Perfection is a fleeting moment. Humans are fallible and so is AI. Trying to sell it in any other form will disappoint yet again.
My statement has nothing to do with not being able to see beyond humanity. I don’t believe in Terminator type problems because as AI evolves I think it will show how well designed the human brain is. Nature and evolution is quite ingenious.
Historical Analysis Is Not Always Correct
Consider the following paragraph from the article.
Technology has upended the financial services industry before. Just think of automated teller machines. Michael Thiemann, CEO of San Diego-based hedge fund firm Investment Science Corp., likens traditional Wall Street traders to personal loan officers at U.S. banks back in the ’80s. Many of these loan officers lost their jobs when banks began assigning scores to customers based on a statistical analysis of their credit histories. In the U.S., those are known as FICO scores, after Minneapolis-based Fair Isaac Corp., which developed them.
I actually have an ax to grind with FICO scores. My wife and I don’t life in the US (Canada/Switzerland), but I take example FICO tests to see how we would rate. The end result is that we are at the bottom of the low risk spectrum. Why? Because we don’t overspend, have no debt and own a single credit card with a honken high limit that we never use.
I know why the example tests (Canadian scores) are giving us a less optimal rating, because we have too few historical data points to give them a picture of what our risk is. Assessing risk based on historical information is a popular endeavor because life tends to follows a Gaussian distribution. The problem with Gaussian distributions is that they are too simplistic and the “rare” events are not as rare as one believes they are. A better assessment would be to use a Gaussian with fat tails.
Even though I have my misgivings about FICO let’s for the moment assume it is a good idea. Does it make sense in the context of trading? No, because there are too many fuzzy data points. FICO works because they have a large enough statistical sampling that the scores are reasonable. The problem of my wife and I is that we are not 100% percent part of the Gaussian curve and to improve our score we need to fit ourselves into the Gaussian curve.
The trading houses want to use AI to answer questions such as the following.
“We want to be able to ask a computer, `Tell me about the merger of corporation A and corporation B,’ or `Tell me about the impact on the markets of sending more troops to Iraq,”’ McKeown, 52, says.
Let’s say we use historical information about merger’s. How many mergers have their been? A few dozen involving publicly traded companies, maybe a few hundred? And of those few hundred how well can they be documented? In contrast Fico probably has close to a 150 million data points and whatever they assess is reasonable.
As one person who was well versed in statistics said to me. Statistics do not predict the future, they are a snapshot of what has occurred in the past. To get an estimate of the future you to employ analysis of variance.
Patterns and Behaviors Don’t Mix
The following statement, while seeming logical is a mistake:
As an exercise, Kearns and his colleagues at Lehman Brothers used such programs to examine orders and improve how the firm executes trades, he says. The programs scanned bids, offers, specific prices and buy and sell orders to find patterns in volatility and prices, he says. Using this information, they taught a computer how to determine the most cost-effective trades.
In the scientific method you are taught to define a process that can be reproduced thus building a theory. In essence you try to find patterns that causes a theory to be constructed. Though the market is not scientific, it is behavioral and I question the effectiveness of patterns.
I prefer to think of the market as a team sport like hockey. I could give an example right now. The Buffalo Sabers are playing against the Ottawa Senator (go sens go) and the Senators are 3 games up. The Sabers have not yet won a game and one player said, “the Senators are doing an incredible job shutting down our powerplay.” The killing of the powerplay was recognized by the Senators as a weakness and they worked doubly hard to overcome the weakness with effective results. So the pattern of inability to kill the powerplay is gone. The Sabers now need to quickly figure out what to do.
My point is that the AI systems are not acting in isolation, they are part of a system that reacts and changes gameplans. A rock when being hit by hammer stays put, not a market.
Why AI and Humans are Fallible
Humans and AI have a common denominator and that is to absorb and process information. Many of us expect that computers with their ability to absorb huge amounts of data can process the data faster than us humans and thus will be able to make decisions we cannot.
Say you have a 500 GB harddisk, and you filled it with pictures that you have taken. If asked to find a picture how long would it take? A few seconds maybe? Not change 50 GB to 40 Terabyte (40000 GB). How long would it take to find a particular picture that you took? Probably too long to make it worth it to keep 40 Terabytes of data. The thinking is that you cannot find the picture because we are limited by the sizes of our brain. We think the problem is that we cannot keep track of all those pictures and thus AI can.
The problem is not the sizes of our brain, but the fact that there is too much information. Here is where I think people are not understanding the problem. AI cannot figure out something that we cannot. Let me illustrate it another way. Say you are trading and you using two indicators. Your trading is doing pretty well, and but then to improve your trading you say, “I need twenty indicators.” Will that improve your trading? You might be tempted to say no because it is too much information. BUT with AI my trading would be better. Wrong answer because with more information does not come better decisions.
What needs to be understood is that the real problem that is being avoided is how to take a large amount of information and distill it such that you can make sense of it. AI cannot magically come up with a magic formula.
Where I can see an advantage with AI is that it can be used to make consistent decisions in domains that we understand, but want to optimize. AI can be used to make trades, but only insofar that the strategy and data structures have been defined. To have AI come up with knowledge out of nothing is stretching the abilities of AI.
Some even dream of having AI scour the Internet in search of nuggets of information. This seems interesting, but I think will turn into fools gold. Think about this. You are using AI to scour the Internet to find a hot tip, or are using the collective of the Internet to come up with a hot tip. Would you trade such a hot tip? If you would, I have some land for you in Florida!