I have - again - investigated the use of neural networks for forex predictions and have not found a really reliable program or way to use it. Users of all the wellknown programs like Brainmaker, neuroshell, Future Wave, C-Trader, Netlab, BioComp, Matlab, NeuroDimensions and others have spoken about and written of in some cases extensive testing with meagre results. An expert who had been on a team in China wrote, they have given up the project. A hedge fund in Berlin invested lots of funds in nn projects and also gave up in the end. Has anyone heard of a positive and continuously reliable use of nn? Thanks for your inputs. Felix
1. NN. is a multiple layer optimization algorithm. As long as you don't go overboard, it's a very powerful tool. 2. If a human can't do it, the computer won't be able to. It's only a computer and they can only follow your instructions to programmed. 3. Restudy Machine learning and AI algorithms. What you mention is outdated and lacks study. It's common knowledge that AI is effectively implemented within a delegated task in an expert system.
NNs is just another method of fitting data to an equation or systems of equations. Of course, it is more involved and complicated than that but this is more or less the end result. It boils down to this question: is there an equation or process that can predict market moves with high probability? Your answer and my answer is as good as everybody's. Nobody knows the answer, IMO. Trading System Lab and Automatic Pattern Search are two programs along these lines that use different techniques to discover trading systems from historical data. The first uses genetic programming with various statistical filters to avoid optimization and curve fitting. I don't know whether the problem is solved and to what degree. The second avoids curve fitting and optimization by simply focussing on price action and as a result it is limited to price patterns. In this case there is potential of selection bias but one can possibly figure out ways to minimize it. These programs I mentioned are white-boxes in the sense that they generate code for various trading platforms. Free Neural Network framework to create, train and test artificial neural net: jooneworld.com
Among the quant traders that I have worked with, none are using neural networks in creating their models/systems. I have also met a number of hedge fund managers in the quant trading space who told me that neural networks are being avoided because none of them have come up with models that work consistently out-of-sample. Interestingly, one of them is a former professor in A.I. who has published a number of rather well-know publications on neural networks. His fund is very successful and he is considered a rising star in the relative value commodity space. He sort of confided to me and my colleagues that the methods he's using are actually quite mundane and straigthforward, and have little in common with the A.I. research he did in the past.
You are correct that NN's don't predict market movement very well. As intradaybill states, genetic programming is more ideal for creating trading models. You seem to mention all of the mainstream and well known AI software out there, but what about a program like Discipulus, which is used as the base software for Trading System Lab? Imagine a computer that programs itself, or a trading system that fits simple rules to a particular market on its own. Since the rules are boolean, curve fitting is less of a problem than it is with neural networks. Returns of 15-30% using this type of model are not impossible.
Interesting. In my fist attempt to make an automated trading system, I actually implemented the algorithm described in one of his papers: http://www.icsi.berkeley.edu/~moody/MoodySaffellTNN01.pdf We did not have much success in reproducing his results though (but it was part of a university project I did, so focus was more on getting the report done, than actually achieving a working system). However, I see that he apparently is doing pretty well: http://www.advancedtrading.com/showArticle.jhtml?articleID=205207957
As a user of a neural-type network I have noticed that most people are not aware of a mathematical formula for relating the size of a sample to the number of inputs. If the inputs increase beyond a few, the sample size required becomes massive -- if the sample size is insufficient then curve fitting is almost guaranteed. "Give me four parameters, and I can fit an elephant. Give me five, and I can wiggle its trunk". --- John Von Neumann