I attended the Missouri University of Science and Technology again, where I received my M.S. in computer science in 2008. My thesis advisor was Dr. Daniel Tauritz.
You can look at the final version of my thesis as a PostScript or as a PDF file. My thesis was an attempt to predict the stock market using leaning classifier systems, a sort of evolutionary alorithm. I called my approach TSC, and it actually produces decent results, but it is not capable of returning at the level I would need to put it into practical use in trading.
You can also look at the slides for my thesis defense from November 2, AD 2007 as a PostScript or as a PDF file.
I gave a lecture for Dr. Tauritz's Computer Science 448 class on October 4, AD 2006, where I outlined my work, as well as giving a basic introduction to reinforcement learning, ZCS, XCS, XCSR, and time series. You can read the lecture notes in either DVI or PDF format.
TSC as presented in my thesis underperformed the Dow Jones Industrial Average. This effectively makes it, in the form presented in the thesis, unusable for real-world application. There are several reasons for this, some of them EA-based, but mostly problem-specific. The most fundamental problem with the system presented in the thesis is the lack of earnings information. This is a major error on my part; the second-most influential element of a stock's price is the underlying company's earnings, and this information is not presented to the system.
You can download the historical data set for the Dow Jones Industrial Average that I used in my thesis as a CSV. This data was downloaded from Yahoo! Finance on August 18, AD 2006. I wouldn't recommend using it though, or any of the common indices for that matter, if you are attempting to use it as a data set for analysis. One of the reasons for reduced performance, which I didn't realize would be an issue at the time, is that the DJIA is not a contiguous index. It is routinely adjusted, with some companies added and others removed, and with the weights altered by Dow Jones and Company; you can read about the historical components here. I would instead recommend using a single issue, or if you insist on using a basket, use one of your own making that doesn't get adjusted for the duration of your study. Below is what the data set looks like, from a monthly perspective.