Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!das-news.harvard.edu!news2.near.net!MathWorks.Com!europa.eng.gtefsd.com!howland.reston.ans.net!pipex!doc.ic.ac.uk!cc.ic.ac.uk!atae
From: atae@spva.ph.ic.ac.uk (Ata Etemadi)
Subject: Re: stock prediction :(
Message-ID: <1994Oct9.002814.5003@cc.ic.ac.uk>
Nntp-Posting-Host: icmag1.sp.ph
Reply-To: atae@spva.ph.ic.ac.uk
Organization: Imperial College of Science, Technology, and Medicine, London, England
References: <CwyKp7.66q@watdragon.uwaterloo.ca> <36i6sb$8j0@newsbf01.news.aol.com>
Date: Sun, 9 Oct 94 00:28:14 BST
Lines: 39

G'day folks

I missed the start of this thread so I hope what I have to say is 
not a repeat. IMHO the real question has nothing to do with ANNs etc.. 
These are the possible solutions and not the problem. One formulation
of the problem is:

  How can we decide if a data set is capable of self-prediction ?

Let me qualify this. Lets say I have a system with N parameters 
(including time) only a fraction F of which are measurable. Some of 
these control parameters may be random and the numbers N and F change 
with time. I also know that there are local and global inter-relations, 
but I do not know their exact form. Can I predict the state of the 
system (or some fraction of it) at time t+1 given its state at time t ? 
I use the term 'predict' loosely in that it may be just to determine a 
trend. Its very important however that I DO KNOW that the effect of what 
I can not measure IS somehow included in the values that I can measure. 
In reality therefore, it is because of this that I will be able to say 
the system IS indeed capable of some degree of self-prediction and this
is also why the ANN approach is to some extent successful in applications 
such as stock markets (specially given the feed-back loop).

The current approach is basically to assume that the answer to the 
original question is 'yes' and start to apply a solution. If your 
solution is successful then ofcourse you can say your original assumption 
was valid. Thats an endemic (or is it epidemic :-) engineering approach, 
but it leaves much to be desired. First of all, you have to do better 
than random and also be sure that a plain linear (or higher order) 
extrapolation is not just as good. Most importantly however, other than 
trying to include more parameters, you don't know how to improve the 
performance of your predictor. All you really have is a black box the 
workings of which are a mystery. I think for anyone really into the 
subject I have said enough to see what is the "formal" approach. After 
some rather bad experiences with the honesty of people from this group 
I think I will leave it at that.

	adios
		Ata <(|)>.
