Some people do, check attachment. If your a believer in Ray Kurzweil's work then NN capabilities should be expanding exponentially and potentially become superior to your NN between the ears within our lifetime if not the next decade.
This is a very poorly done study. In any case, this is an angle that has been pretty well explored. The "feed-everything to an optimizer and let it work out a trading strategy" route is pretty well treaded at this point, with little result.
Some articles on neural nets in financial markets are below. What I'm curious about is since neural nets are universal function generators, shouldn't they be able to do anything regular statistics can do if as a previous poster said, you have the right inputs and outputs? I'm actually working on designing a neural net from scratch so hopefully everyone's wrong about them having no chance to work They'reall 1st author, et. al. and all are available online in PDF format for free: An Empirical Analysis of Data Requirements for Financial Forecasting with Neural Networks STEVEN WALCZAK Applying Fundamental Analysis and Neural Networks in the Australian Stockmarket Bruce Vanstone, et al ENHANCING SECURITY SELECTION IN THE AUSTRALIAN STOCKMARKET USING FUNDAMENTAL ANALYSIS AND NEURAL NETWORKS Bruce Vanstone A neural network approach to futures trading Hong Pi, et. al Evolving Neural Networks for Hang Seng Stock Index Forecast Yong Liu Financial Forecasting through Unsupervised Clustering and Neural Networks N.G. Pavlidis "Forecasting the 30-year U.S. Treasury Bond with a System of Neural Networks" Wei Cheng Neural Networks for technical analysis: A study on KLCI Jingtao Yao Recent Developments of Self-Organising Modeling in Prediction and Analysis of Stock Market Ivakhnenko
Given: 1. infinite computational power 2. an infinite number of independent observations (data) Yes. But under any realistic assumptions, No. The problem is the rapid (exponential?) expansion in the amount of data required to train, without overfitting, an increasing number of model parameters -- here, artificial neurons.
The user does not need to understand how these NN programs work to use them. All that is required to use them correctly is to follow closely the instructions of the developer. One problem is that most people get these programs and never read the manual, they treat them like computer games and try to learn how to use them in a trail-and-error mode. Others want to know how they work so that they can reverse engineer some of those products. I think there are some of this type in this forum. I use a data mining program of the algorithmic type (not an NN I am sure) and when I talked to a number of other users of it in the past I realized they had not even read the manual. One must invest time and effort to use a tool properly. It is the wrong approach to try to understand how and why it does what it does. The developer might have spent years working on it after spending years in college learning a specialized subject. If you do not have a Ph.d in AI (and I don't) when you read some of the papers you probably draw the wrong conclusions about things you don't even understand what they mean.
To be a "universal function generator", you need neurons => infinity which roughly implies examples => 2 ^ infinity == infinity. So the relative suitability of ANN's really depends on how a LESS THAN UNIVERSAL neural net learns the "true function" you're trying to model, versus some other method (whether it be statistics-based or otherwise). So you need to pick your model/method well, and hope/verify/compute whether you're not fitting too many parameters.
How would we know the result if successful people were keeping it to themselves? If I discover a trading method by any means whatsoever, that is scalable and highly profitable... I have a choice of putting in for some bullshit prize and divulging what I know and making it available to every rapacious sick bastard on the planet.... or making some huge bucks and trampling those same rapacious bastards... gee, which is most satisfying...
The goal is most definitely not to learn a 'true' function. You will never know the 'true' function, regardless of how many neurons you have, unless you can some how time travel to the future and have a database of all past/future data; not only is that unnecessary, but it is counter productive. The goal is to find some kind of relationship between the three criteria I mentioned, that nets you a positive expectation over time. If you really want to convince yourself of its effectiveness as a tool: 1) Do some research to understand the basics of what it is meant to do. 2) Read up on some academic papers to get a feel for what has been attempted and where progress is being made. Regardless of what anyone tells you, there are quite a few bright academics around. Although it certainly pays to be skeptical if you see they sampled a window of 1 year to draw conclusions (kind of like how uptick rule study was done). 3) If you feel you have a rudimentary understanding, there is no better teacher than to build the model yourself and prove to yourself how useful it may or may not be. And a bit of advice, go back and look at what the most difficult part is-- steps 1->3, NOT so much the architecture. If you think a NN is complicated, go back to a simple linear regression and try to understand it's usefulness and/or pitfalls. You'll likely see how the 3 steps are much more important, than the learner. Maybe there are advantages that a NN has over simple regression? Seeing the truth for yourself is miles ahead of what anyone can tell you. And it is quite easy to build with modern tools that cost you zero. Unfortunately, that is the problem with simply reading 'papers,' they tend to be a bit skimpy on how to actually replicate or build on their work.
dtrader98, I agree with virtually everything in your post; in fact, I think it in itself is quite a good primer for anyone wanting to apply any "computer fitted model" to trading. But it bears only an indirect relationship with the discussion I was having with the earlier poster, whose statements implied the tractability of using ANN's as a "universal function generator", which by definition implies the goal of learning a 'true' function. Irrespective of that, I think all the readers here would do well to heed the methodology outlined in your post. Especially before shelling out a lot of money for a commercial "holy grail" product.