[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: MKT: Trendy Neural Net



PureBytes Links

Trading Reference Links

Allan P. Harris wrote:
> 
       We have gone to great extremes at the web site to make it clear
that we do
> not offer trading advice.  We offer simulated models as examples of how the
> neural net can be integrated into a mechanical trading campaign.  We
> thought that by clearly describing the entry and exit parameters for these
> mechanical models and faithfully tracking the hypothetical trades based on
> said parameters, that the published results would reflect more clearly upon
> the effectiveness of the neural net then upon the vagaries of individual
> entries and exits.


> 
> >How can you assume that the price you are getting at 10 pm is the same
> >as the price was at 4:15 (entries) or the same as the next day;s open
> >(exits).
> 
>         This assumption was not pulled out of thin air. Prior to establishing and
> publishing the mechanical models, we studied the price action of Globex
> between 7:00PM and 10:00PM ET as it related to the day session closing
> prices.  What we found was that the closing price of the day session
> usually available during the evening session.  When it was not, we found
> that it was just as likely that you would get a better price, LONG or
> SHORT, then a worse price then the close of the day session.  Accordingly,
> we thought that using the closing price if the day session was a fair
> representation of system performance.  Not perfect, just fair.  Recently we
> have seen huge moves start in the evening session, about the time the
> foreign markets are opening.  Again, we are finding that the huge moves are
> resulting in an equal number of "better' versus "worse" fills from the day
> session close.

I realize this disclaimer is so you don't have to register as a CTA,
that is fine, but you are selling signal and I am just trying to get
more info about the system so that I can make an informed decision about
it myself.  And I commend you on mentioning the assumptions you make
several times on your website.  You are certainly being upfront about
it.  I just get bothered with any track record that isn't real time and
when you specifically mentioned that you were short for Friday Nov 7 it
made me stop and think.  If I am not mistaken, by 10 pm on Thurs night ,
I believe the SPoos were locked limit down from the 4:15 close so your
short was impossible to enter let alone enter at the price assumed.  I
know some fills will be better and some worse, but it is just bothersome
to me.  That is the limitation of simulations.

> 
> >Surely you could afford to trade a mini to get an actual real
> >time track record. No??
> 
>         We are and are quite pleased with the results.  We will soon be publishing
> the Trendy Systems, LLC actual futures' account statements at the web site.
>  We were hoping to have at least six months of trading the account in
> before boasting, which would be in January of 1998. There are also some
> legal issues that we think are minor and can be resolved.

Excellent.

> 
> >I also wonder if you could give us a quick rundown of why your neural
> >net is different from others.
> 
>         We don't know what it is in the analysis of a dozen inputs, looked at in
> hundreds of different arrays, that the neural net is seeing that causes it
> to generate it's output from which we are gleaning Buy and Sell decisions.
> So it is objectively impossible to answer your question.  

The 18 month real time track record, simulated or not, may offer some
tangential evidence of
> effectiveness that distinguishes Trendy from other neural nets. 

This is the trouble, a real time track record can't (by definition) be
simulated.  A track record is either real or hypothetical by CFTC ruling
and simulated is hypotheitcal.

 Or it may
> be just a random occurrence of accuracy that is now on the cusp of failure
> and then oblivion, like all of the others you mention. For the time being,
> we are content that it is reporting something of predictive value, which in
> and of itself makes it different from other neural nets.
> 
> Allan P. Harris
> http://www.TrendySystems.com

Again, I am just trying to get a better understanding of what neural
nets are.  Intuitively, they seem like they have potential in trading. 
But as Walt's post indicates some big money has been spent by some
fairly smart people to develop them and they have generally failed.  Any
opinion on why they failed (if that info is correct)?  Anything you can
give, rather than... it just works, but we don't know exactly how...
would help me understand a bit better.

Thanks,
Eric