|
PureBytes Links
Trading Reference Links
|
Dans un courrier daté du 17/03/99 07:08:25 Heure d7iver Pari25 Madrid,
marlowec@xxxxxxx a écrit :
> developed a DLL that is a fuzzy logic engine. It works off a truth
> table of two independent variables. I really haven't perfected it
> into a trading system, but rather as a FL test bed. My original
> version was all in EasyLanguage code linked together via Global DLL.
> It ran quite slow and the true DLL version was a major improvement.
> I'm a bit reluctant to publish the code, but maybe in the future I
> will share it with this group. FYI, Earle Cox book on Fuzzy Logic
> Systems Engineering has all the info one needs to build a FL
> controller or trading system.
>
> While I'm at it let me add that Fuzzy Logic is not the Holy Grail.
> It is a tool for combining a lot of disparate factors to arrive at an
> answer. In other words you can intelligently add up a variety of
> conflicting signals to yield a composite signal. In my case I was
> looking at the rate of change of price peaks and the rate of change
> of price valleys to generate a composite trading signal. I believe
> it yielded good entry points (but exits stunk).
>
> I got off on a tangent to optimize the truth table (25 variables).
> Talk about over optimizing! Beautiful results until tested out of
> sample.
>
As for anything, a technique is nothing else than a technique.
The key to the problem is the way to apply it .
If you attempt tooptimize a truth able for 25 variable, you will have more
rules (cells) than the cases you may expecte to see in a very large datbase.
Building such systems requires to clever control the degrees of freedom.
That's aside, it is not necessary to use 25 variables,most of them will show
anannoying redundancy.
An other problem with fuzzy logic is thta it is difficult to optimise due to
the numerous parameters ( fuzzy sets position, shapes) and control what
optimisation really does.
We have chosen the fuzzy logic + neural network solution because we are able
to minimize the drawback of both techniques ( fuzzy sets comlplexity is
greatly reduced by using neural networks interpretation of the fuzzy rules,
and NN are not so easily allowed to do their usual overfitting because they
need to stick to the strict interpretation of fuzzy rules,limiting this to the
necessary cases).
This is the reason why the neurofuzzy logic approach yields to results that
are generally correct on unseen data..
Those who may not believe it can download the free evaluatio version from the
web and do their own experiment with their own data ( Tradetation templates
are also provided).
Sincerely,
-Pierre Orphelin
www.sirtrade.com
|