|
PureBytes Links
Trading Reference Links
|
> -----Message d'origine-----
> De : Gary Fritz [mailto:fritz@xxxxxxxx]
> Envoye : dimanche 17 mars 2002 19:13
> A : sirtrade@xxxxxxxxxx
> Cc : omega-list@xxxxxxxxxx
> Objet : RE: A Mythomaniac Professional comments on Pierre's software - -
> -
>
>
> > Complexity will never turn anything that fairly works into a
> > miracle trading system, but will enhance things that cannot be
> > otherwise: Removing some bad trades, displace some of thems fom one
> > or two bar backward, reduce the max DD.
>
> Pierre, something I've never understood about this approach:
>
> How can you adjust a trade there, displace an entry here, etc,
> without falling into a curve-fitting trap? You say that Safir-X
> generates thousands of rules internally. That sounds to me like a
> perfect setup for curve-fitting.
This is not the case because the rules are established first on a bar by bar
basis ( not on a pattern basis), means that the rules are diffferent if the
underlying indicators move from a small amount.
The software do not even know what is a pattern or a trade to identify. If it
has 4 inputs,it has only 4 numbers to mix inorder to give a local answer that
is +1 or 1.One may speak of micropattern if you want, with only 4 values at
bar N and that is allof what the software see.
This is for currentbar only, and change every bar.
But they are treated all toghether in the decision tree rule, which is validated
agains the whole serie of data, this time regardless what the micro pattern
could have been at bar N. I mean that when the decision to keep a solution ( a
decison tree) is taken, it is globally done ,only by evaluating the performance
summary.
So there are two concurential examination of the data by Safir-X: A local way
where the answer is judged from it's local pertinence, regardless of what the
other rules at other bars may do, and globally, with one field of the
performance summary, where the information of the pertinence of each rule is
unknown for the final decision.
This is in fact more complex in real safir-X world, but the two process beign
judged independently, there is no direct action between the local curve fitting
and the global curve fitting that acts at a higher global level without
retroaction on the rules.
>
> I know you say these rules generalize well on unseen data, but....
> how?? How can they, when you have more rules than you have trades in
> the training period!?
A Safir-X rule is a per bar rule. But even if you have thousands of rules, most
of them are fired in exceptional cases, that are due to the exterme values of
some indicators or quite impossible case ( for example, you will never have to
fire "if RSI(c,10) very oversold and SLOWK(12) very overbought then..." Most of
the time, the system fires the same 50 -100 rules, frequently, and the others
by exception if case apply.
Due to the fact that the frequency of the rules is determined not by the
frequence of the trading patterns ( there are no patterns) but by the frequency
of the simlar cases that have been identified by the fuzzy engine, and knowing
that the global selection has been applied without the hinsight of what rules
were locally pertinent or not, this dichotomy in the process yields to system
that have a lot of rules, where 90 % are fired 10% of the time when 10% are
fired 90% of the time ( this is an example), where their local value is
independent of their global value ( here you may speak of existing patterns,
that are the trades producing the performance summary).
This means that it is very difficult to overfit this kind of model, and why we
have to discard the classical objection that if the number of rules is of the
same order of magnitude of the number of trades, the overfiting issue is of
question.
In this case, it is not, because of the unequal distribution of firing occurence
that center on a few dozen of rules, regardless of the database used for
training.
This also means that if an indicator set is not well suited to trading, the
software does not find anything of value, and if by chance it worked, it will
fail on useen data without any doubt ( this behaviour ofthe software may have
hurt some psychiatrist in the past)
Well, it'es easier for me to explain in french with a paper and pencil
>
> Gary
Sincerely,
Pierre Orphelin
www.sirtrade.com
TradeStation Technologies representative in France
Safir-X neurofuzzy logic trading system builder
To subscribe to our FREE trading system newsletter:
http://www.sirtrade.com/newsletter.htm
>
|