|
PureBytes Links
Trading Reference Links
|
Mark,
as "indicators" I used adaptive neural nets (subnets which retrain
themselves every day on a certain "lookback" period). These subnets on their
part can use as their inputs different indicators whose behaviors they try
to "learn". Neuroshell has more than 800 predefined indicators ranging from
price data to Fourier analysis and wavelets. You just take a handful that
seem plausible to you and see what happens. The actual neural net I designed
then learned the behavior of the subnets which in turn learn the behavior of
the chosen indicators which in turn are supposed to be representative of
price behavior.
As to the sample size, don't forget these are 30-min. bars, so three and a
half months is a comparatively long testing period. A training set of 10
months was used for that net; I believe it doesn't make much sense to train
a neural net over extended periods of time because it may be learning
behaviors that are no longer valid in the current market period. Each
half-year period would probably require a separate, completely new net which
would have to be trained on current data. In this respect, the rules that
apply to static trading systems a la Tradestation do not seem to be fully
applicable to neural nets.
But I am still quite new at this game and therefore reserve the right to eat
my words at any time without warning. :-)
Best,
Michael Suesserott
> -----Ursprungliche Nachricht-----
> Von: M. Simms [mailto:prosys@xxxxxxxxxxxxxxxx]
> Gesendet: Thursday, January 31, 2002 23:51
> An: realtraders@xxxxxxxxxxxxxxx
> Betreff: RE: [RT] Neural Nets
>
>
> Mike - important to know exactly WHAT did you feed the system ?
> stochastics, trend definitions, etc, etc. ???
> Results look good, but the sample size is a bit small.
> Can you run it thru the S&P 500 - 1997 to 2001 ?
>
>
> > -----Original Message-----
> > From: MikeSuesserott@xxxxxxxxxxx [mailto:MikeSuesserott@xxxxxxxxxxx]
> > Sent: Thursday, January 31, 2002 4:53 PM
> > To: omega-list@xxxxxxxxxx; realtraders@xxxxxxxxxxxxxxx
> > Subject: [RT] Neural Nets
> >
> >
> > Have been playing around a little with a neat present I had bought for
> > myself last Christmas - Neuroshell Daytrader
> > (www.neuroshell.com). Here are
> > the results of one system that came up in testing. As you will
> > see from the
> > attached chart, neural nets are by no means perfect, but it looks
> > like they
> > can make money for you.
> >
> > This very crude QQQ system is based on 30-min. data starting
> Jan. 2001. It
> > doesn't use stops and is always in the market. The net was
> trained on data
> > up to Oct. 9, 2001, then tested on out-of-sample data from Oct.
> 10 through
> > Oct. 24. Following that, it was retrained using the additional
> > two weeks of
> > data up to Oct. 24, and applied on the out-of-sample data from Oct. 25
> > through Nov. 7. In this manner, a series of 2-week-long
> walk-forward tests
> > on out-of-sample data was produced, with the net being
> retrained every two
> > weeks the way it would have been done in real trading, too.
> >
> > In spite of some whipsaw, the system proves quite profitable with an
> > annualized return of 193.9% (statistics attached). While I still reserve
> > judgment on the general value and applicability for my trading,
> I must say
> > the results of testing so far look promising.
> >
> > Best regards,
> >
> > Michael Suesserott
> >
> > Disclaimer: I have no connections with Ward Systems except as a
> customer.
> >
> >
> > To unsubscribe from this group, send an email to:
> > realtraders-unsubscribe@xxxxxxxxxxxxxxx
> >
> >
> >
> > Your use of Yahoo! Groups is subject to
http://docs.yahoo.com/info/terms/
>
>
To unsubscribe from this group, send an email to:
realtraders-unsubscribe@xxxxxxxxxxxxxxx
Your use of Yahoo! Groups is subject to http://docs.yahoo.com/info/terms/
------------------------ Yahoo! Groups Sponsor ---------------------~-->
Access Your PC from Anywhere
It's Easy. It's Fun. - Free Download
http://us.click.yahoo.com/xnjxPB/0akDAA/cosFAA/zMEolB/TM
---------------------------------------------------------------------~->
To unsubscribe from this group, send an email to:
realtraders-unsubscribe@xxxxxxxxxxxxxxx
Your use of Yahoo! Groups is subject to http://docs.yahoo.com/info/terms/
|