|
PureBytes Links
Trading Reference Links
|
Hate to say this, but I agree with Pierre..
having spent the best part of 5 years full time working with nnets, using them to predict markets (among other things), Pierre is exactly on the money.
Getting them to behave and produce useful results is definitely non trivial.
The hardest thing is asking them the right question. Second hardest is munging the data going in and out to an 'interpretable' format.
Oh, and the biggest risk is over optimisation. If you think about it - with, say, a typical 3 layer backprop net, that has x input nodes, y in the middle layer, and z output nodes.. you're talking x * y * z degrees of freedom.. now imagine coming up with a quadratic equation with x * y * z number of elements, and you can imagine how curve fitted that would be.
Essentially, nnets tend to result in a blurred average.. something which is done much simpler and more reliably with an -actual- smoothed average, ala T3 or the recent conversation by Bob, Bob, Alex et al.
By the time you figure out an appropriate question to ask, and how to formulate the information you're passing in, I'd bet anything you can find a more transparent, comprehendable, reliable method of achieving the answer.
Oh, and as for Ward Systems Group, yeah, they've got some good stuff.. I used to work with them.. in fact, developed the first version of what would have been their NeuroShell Trader.
Just my 5c.
Si
At 05:23 27/06/2004, Pierre Orphelin wrote:
>Training a NN for a supposed good result is not a piece of cake. Even if you are
>an expert.
>The problem has too many solutions and false tracks to follow and you may spend
>a lifetime doing such a research .
>Think about the number of possibilities between various architectures and
>target of fitness functions.
>
>An alternate solution is to use a 100 % automated research tool that will do
>the expert work for you.
>Safir-X is one of the few of them.
>You may ask a free evalauation version. Cheaper than hiring an expert and
>immediate results granted.
>
>Sincerely,
>
>Pierre Orphelin
>www.sirtrade.com
>Tradestation 2000i, TradeStation 8 sales and support
>Safir-X, neurofuzzy logic trading system builder
>
>To subscribe to our FREE trading system newsletter:
>http://www.sirtrade.com/newsletter.htm
>
>
>
>-----Message d'origine-----
>De : Allan Kaminsky [mailto:allan@xxxxxxxx]
>Envoye : samedi 26 juin 2004 15:00
>A : 'robert pisani'; omega-list@xxxxxxxxxx
>Objet : RE: neural net expert
>
>
>You might check out Ward Systems Group.
>
>They make superb, innovative NN products and know finance (one of their
>products is the incredible Neuroshell Trader).
>
>They offer a service that might be just what you're looking for. See:
>
>http://www.formuladetective.com/
>
>Allan
>
>-----Original Message-----
>From: robert pisani [mailto:r.pisani@xxxxxxx]
>Sent: Saturday, June 26, 2004 12:31 AM
>To: omega-list@xxxxxxxxxx
>Cc: realtraders@xxxxxxxxxxx
>Subject: neural net expert
>
>I have a data base of inputs and a number of target functions. It is clear
>to me from technical considerations that neural nets can be developed to
>estimate these target functions. I want to do this in some optimal
>fashion. I am well-informed and technically competent, but by no means
>expert in neural nets and their numerous architectures, and I would like to
>hire someone to develop low rmse trained neural nets for me, and perhaps to
>educate me in the art of developing them. The neural net expert does not
>have to know anything about markets.
>
>Any ideas?
|