Ask HN: Has anyone used machine learning to trade?

I do this and I make money in 1 out 2 trades. Do you know what the heck you are talking about ?

http://www.brettsteenbarger.com/articles.htm

How much do you know about trading now? If you know little, I would recommend the textbook Trading & Exchanges by Larry Harris. It's somewhat out of date but still very good. I work at a high-frequency trading shop and that is the first book they shove at you when you walk in the door.

Other than that, I'd like to second what several other people have already said: getting a job at a firm is a good way to learn, you're outgunned in terms of brains and money, and 10k is probably too little. Now, you can do a lot of learning for 0k while you save money, and every startup is outgunned by goog, msft, but that doesn't mean there isn't room in the ecosystem. Like a startup against an established player, you could find a niche that is going unexploited. What makes trading harder is the established players don't tell you much about what they're doing, so you don't know if you are doing something new or walking into a trading strategy perfected 10 years ago by DE Shaw. With a software startup, you have more information about what your competitors are selling, what their products are, how you might differentiate yourself, what's missing from the internet, etc. Software companies promote their products and try to get public attention; trading shops try to keep as many secrets as possible.

You might find it interesting to look at this site, which gives you a lot of the infrastructure for testing your FX strategies with real time data (although not direct access to the data).

http://preview.collective2.com/

Strategies are independently monitored and ranked on the site, and can be switched from simulations to real accounts when you're ready. If your strategy performs well, it will be noticed and you can get paid commissions on it (potentially helpful for raising capital).

For FX and equity data, you might as well use Yahoo finance. I run a Slicehost server that scrapes it once every minute with no problem. It's not ultra high frequency but you can't hope to compete in that space anyway, as others have pointed out. You should start building a historical database now before you do anything else, so that you'll have something ready when you need it. You probably don't need years and years of it.

I worked as a quant for a while and would support the other comments that standard ML techniques are unlikely to outperform, but you might be the exception.

Please read http://en.wikipedia.org/wiki/Fooled_by_Randomness before you take a plunge.

All I can say is that 8/10 are losers is highly improbable. Given no other information, if you enter a trade there is a 50/50 chance that the next tick is going to be up or down. But notice I said given no other information, there is much more information from the behavior of the charts. If you are a small player the charts will help you be right more than 50% of the time. The rest is how your algorithms and trading methodology take advantage of the information.