Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!satisfied.elf.com!news.mathworks.com!news.duke.edu!godot.cc.duq.edu!toads.pgh.pa.us!investor.pgh.pa.us!rbp
From: rbp@investor.pgh.pa.us (Bob Peirce #305)
Subject: Re: Neural-nets and trading
Message-ID: <1995Jan30.191035.2421@investor.pgh.pa.us>
Date: Mon, 30 Jan 95 19:10:35 GMT
Reply-To: rbp@investor.pgh.pa.us (Bob Peirce #305)
References: <jaad.1.0012BFEE@knoware.nl> <3gfval$4fm@kitten.umdc.umu.se>
Organization: Cookson, Peirce & Co., Pittsburgh, PA
Lines: 58

In article <3gfval$4fm@kitten.umdc.umu.se> Roger:olsson@jus.umu.se writes:
>> I am looking for people who are using Neural nets and other AI-tools for 
>> trading financial markets.
>> 
>I'm looking into it right now. Have been using an application
>for windows, doing technical analysis with good results.
>I gather its not possible at all to use a neural network
>application to get even better results.

Depends on what you mean by better.  I am very new to neural networks,
but I think they are going to be very useful to prove the validity of
various approaches.

One of the problems I have always had was to go back in history to ask
what do I know today and what can I predict from it?  This would have
taken many months, even years, with my previous methods.  With NN I can
start at a point 30-40 years ago.  Learn what there is to learn and use
that to project the future.  I can then roll foreward in a very short
time to look at each cycle as it evolves.  This would have been
impossible before.

If I run a NN on the entire history, I get remarkably good results, much
better than I ever got before.  However, some of the stuff I "knew,"
say, in 1970 I didn't really "learn" until, say, 1983.  That is total
nonsense.  Unfortunately, that is exactly the way much security analysis
is actually practiced. 

When I run a "look ahead" model the results are much worse than this
ideal.  However, what pleased me was they still were very good, and
probably much better than most people ever have any hope of doing. 

For example, one of the first models I studied returned 15.5% per annum,
in terms of the Total Return S&P, from 12/61 to 12/94.  The Total Return
S&P only grew 9.9%.  Most folks, myself included, would be very happy to
get that kind of return.  When you add in that it is all based on
looking ahead and not depending on things you couldn't possibly know, it
just gets all the better.  For me it proves that it is possible to do it
and the only question becomes to find the best way.

That's where I have the most trouble.  I have been trying to read a lot
of stuff aimed at the layman and what I keep running into is that there
doesn't seem to be much of a theory about how to design the hidden
layers in terms of numbers, nodes and inter-connections.  I have ended
up with 18 models to examine with one and two hidden layers of various
sizes, with and without connections from input to output.  They all
produce different results and I really don't know why.  So, what I am
looking for is consistency from cycle to cycle, indicating a stable
model, or improvement form cycle to cycle, indicating a model that is
better the more it learns.  Needless to say, I don't have either yet,
but it is pretty clear to me I have a better chance with NNs than with
anything else I have used.
-- 
Bob Peirce                Pittsburgh, PA                 412-471-5320
rbp@investor.pgh.pa.us [OFFICE]    me@venetia.pgh.pa.us [HOME (NeXT)]

There is only one basic human right, the right to do as you damn well
please. And with it comes the only basic human duty, the duty to take
the consequences. -- P.J. O'Rourke
