Stock Prediction & Machine Learning

I don't think there is a book specifically focused on machine learning and finance, although there are tons of books on computational finance. A lot of these methods assume stock prices fluctuate with respect to a specific statistical model, then estimate the parameters of the model from past data, then attempt to predict future trends or volatility. I've done work with some of these models only to test out new Monte Carlo methods so I can't really recommend papers for actual market prediction, only ones that are hard to do inference in, which is not likely to be helpful.

There was a NIPS workshop in 2005 it has links to a few authors, John Moody, Ramazan Gençay, Amir Atiya. Might be a good place to get started

As far as I understand, the general academic consensus is that its not possible to make any useful predictions for trading because of market efficiency. If any temporal dependencies do exist they are extremely non-linear and time dependent. Unfortunately, this means that the best model for your data is a random walk.

In general, there's not a lot of published literature that is useful for profitably trading the stock market. There is even less that has anything to do with machine learning.

However, if your interest is purely academic, which to me seems to be the case, Google Scholar is your friend. There are tons of papers. Searching for some obvious terms and following citations should give a good feel for what's out there.

For an example of a recent paper, see "Twitter mood predicts the stock market" by Johan Bollen, Huina Mao, Xiao-Jun Zeng. I have not done any work to verify their result, so have no view on its validity.

This is a current article that is a pretty good place to start with a references to a lot of current docs.

http://www.princeton.edu/~aspremon/svmFinance.pdf

Though most of it reads like a children's book, "The Little Book" by Greenblatt is an excellent practical book about automated investing. The appendix provides the interesting details as well as questions you'd naturally ask about backtesting, such as whether that financial data which was used to make decisions on a certain date was actually available on that date. I also like the fact that the calculations are relatively simple (less likely to overfit) and are insensitive to magnitude (he uses a sum of ranks to determine the "top" stocks).

http://www.magicformulainvesting.com/about_book.html

I was sufficiently impressed that I started using the strategy for my own investing 5 years ago...

Starting in the late 1980s, there was a flurry of writing on this subject which trailed off into the late 1990s. The authors ran the range from relatively serious to "get rich quick". Most of these books and articles were pretty much what you'd expect: 1. obtain historic data, 2. train one or more models and 3. use them to guide investment, but a few of them had some interesting ideas. The books are not hard to find: just search for things like "neural network" AND "stock market", mostly with publication dates in the 1990s. Many such articles were published in the magazines "AI Expert" and "PC AI", if you can find old issues of them.

Another area of writing which might be of interest would be popular finance material on the efficient market hypothesis. This gets kicked around on Internet fora which still discuss the use of machines to invest.