about talos works | historical performance | today’s suggestions

In addition to advanced, custom business programming solutions, Pocket Knife Software founder Rick Cockerham has created an elite initiative – Talos Works – to explore the potential for practical, real-world applications for Artificial Intelligence. A life-long practitioner in the realm of theoretical computer science, Rick's initial Talos Works project is based on his hypothesis that: "There is an equation, albeit very complex, that will tell you when to buy and sell a stock – perfectly."

This work involves writing and revising an extremely complex algorithm that possesses the ability to evolve over time. This type of algorithm is know as a genetic algorithm and uses the techniques of evolutionary biology to improve output results.

Click Here to see how well our AI's are performing in the stock market this year.

Rick describes his project explorations to date as such:

"I wrote the actual genetic algorithm very quickly. It probably took me less than a month. But, when I started it I was surprised when it made a lot of money very quickly. I remember I had one individual that made $640 billion in 20 years on a $10,000 investment.

"After some investigation I learned that the algorithm matched the pattern of one stock perfectly, meaning it knew exactly when to buy and sell one stock, but only during that time period. So, the equation would be useless for any other stock or even the same stock in the future. I learned that measuring only the profit was a mistake.

"After that I added 30 stocks from the Nasdaq 100 and S&P 100. I also created a more sophisticated measurement of its success. I started to keep five individuals from each generation: the best average percent, the best profit, the best vs. buy and hold, the best minimum profit, and the best 'but one.' The best 'but one' involved averaging profit after taking away the algorithm’s best stock. That way it was not allowed to run away with just one stock.

"I ran the program for many weeks, but the algorithm was unable to improve on its averages, so I assumed it found the best possible output. I stopped the program, saved the equations it found and then restarted it just to see if it would settle on the same equation again.

"I was surprised when it came up with a different equation that was better than the first, but it only took about a half a day to find it. I did the math on the number of possible equations. My program was locked into what AI people call a 'local maximum'. I was unable to find a way to ignore the local maximums. It all depended on the algorithm’s random starting point. I decided to cheat and just explore many different random starting points.

"I wrote a control program to run the AI for only about a million generations, save the equations, then start over. It would track the number generations the AI remained stuck on one equation. If it didn't improve the equation for 5,000 generations, it would start over.

"I've been running this final program for almost a year now. It has generated about 100 different equations that are all very good at playing the stock market. I wrote a program to pull the latest stock prices from the Internet and feed them to a database. Every night I run the AI equations against the new data. Each one of them was given $10,000 last July and asked to buy and sell just a few stocks that I'm interested in.

"When I looked at the graphs of their trades it was apparent that each of the AIs had a 'personality'. Some of them make hundreds of trades. Some buy and hold a stock for longer. Some are very good at avoiding downturns in the stock. Some are very good at taking advantage of upturns. Each has a certain type or types of stocks they are good at. Some are good at stocks with low prices and high volatility, while others are good at high price and low volatility.

"Given these variations in the equations, I haven't decided how to actually use this information. On any given day the four best equations I picked will give three different opinions on how to trade a stock.

"It's difficult to decide which one to listen to because they all are making money on that stock. My current hypothesis is you have to listen to only one of the AI’s every day. If you try to make all the trades recommended by all the AI’s, you won't do well.

"I will of course continue to ponder the questions raised and find a solution."

©2008 Pocket Knife Software Inc. All rights reserved.