We commonly get people asking about the use of neural nets with financial prices.
Standard back-propagation neural nets are generally good when the solution space is finite/discrete. Such neural nets are good at recognizing what they have seen in the training set–common uses are in credit scoring and facial recognition.
It’s very easy to over-train a neural net and often difficult to tell if it is over-trained. Neural nets on their own tend not to work well with financial prices because the solution space is continuous and stochastic vs. discrete and finite.
Think of a neural net like a camera, especially when over-trained. Such a neural net could learn to recognize thousands of faces–but fail when upon encountering a face it did not see in its training set.
Statistical/kernel methods are far more effective for financial prices.
There are many other interesting machine learning methods, including decision tress and even some hybrid neural nets. Still, the real problem is always your choice of data representation. Depending on what you want to achieve and with a suitable representation you might even get good results from a basic neural net.
Tags: Machine Learning