I am applying that computational neuroscience math to stocks. I was in the process of running a Fourier decomposition on Starbuck’s data for the past three years (in R) when I realized that my sampling frequency was lower than 1.157*10^-5 Hz (longer than 1 day). While this resolution gives me a good look at the low-pass trends, I am interested in day-trading. So, I decided to find higher resolution data to train and test my really simple feed-forward network on.

After an afternoon of searching through companies offering 10 years of historical stock data for anywhere between 5 euros a stock for 1 month of data to $85,000 for all stocks for 12 years, I finally found what I was looking for. $50 gets you a DVD with bihourly data for every useful US equity for the last year. When it arrives, I plan on cross-validating it with Google Finance to see how accurate it is.

http://www.equitymarketdata.com/historical

You can also get a free 14 day trial: https://www.equitymarketdata.com/login