|
PureBytes Links
Trading Reference Links
|
Peter -
Neural Networks are just a parameterized combination of functions which
use historical data and an optimization algorithm to calculate the
parameters. The functions typically used have been proven to be able to
fit almost any curve.
What this means is that a neural net training algorithm is a curve
fitting algorithm for a very complex curve with up to hundreds of
parameters. Most common algorithms absolutely suffer from the "curse of
dimensionality", which means that if you give the algorithm lots of
correlated data, the quality of the results will suffer. Throwing
everything but the kitchen sink at a neural net is not generally a good
idea.
There are a large variety of techniques available to reduce the problems
caused by correlated data. This field is called principle component
analysis. One method is called Singualr Value Decomposition. See Jurik
Researches website for a product called DDR which uses this algorithm
(www.jurikres.com). I'm not affilited with them, BTW. Another way is to
use a different type of neural network.
Larry Lewis
-----Original Message-----
From: Peter Ryan (reply to: pryan@xxxxxxxxxxxxxx)
[SMTP:pryan@xxxxxxxxxxxxxx (MIME) (actually from:
owner-realtraders@xxxxxxxxxxxxxx)]
Sent: Sunday, October 18, 1998 11:32 AM
To: realtraders
Subject: Neuroshell trader
Is anyone else using Neuroshell trader ?
I have had the software for a while, but only started working with it
seriously after reading the 'fear index' posts started by Walt Downs
posted.
It seems like you can just jam every possible factor at it, and let it
sort
out which are useful.
At least, that's what I have been doing, but with only mixed results so
far.
I would be interested in other opinions.
Thanks
Peter Ryan
|