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

RE: [english 100%] Genetic Programs and Neurofuzzy Logic


  • Date: Mon, 24 Aug 2009 23:14:21 +0200
  • From: "Pierre Orphelin" <pierre.orphelin@xxxxxxx>
  • Subject: RE: [english 100%] Genetic Programs and Neurofuzzy Logic

PureBytes Links

Trading Reference Links

Genetic  algorithms (GA)  are  methods  of evolving   optimization of
rules, parameters.
They speed  up  the  trial and   error  optimization process by  zillions vs
human  coding trial and  erro testing.
In this sense  they are in  the  top  notch optimization  category,  the
drawback being  that they  can  find  solutions  everywhere on  any data
set. 
Of  course,  the  problem soon arise  with  unseen  data,  where  the
dazzling  solutions  miserably  fail, excepted  by  chance.

Neurofuzzy  logic  is a mathematical based  method  of  encoding / decoding
the information,  that  work  better  when  the   raw  info are  not as
crisp as  expected (  market  data are a   good example of  unclear  data,
and  most of   human  activities share  this. In this sense,  fuzzy  logic
is a  privilegiate  way  to  encode / decode  uncertain data  for  further
easier treatment).   

We   do  not  use  GA  methods  to  build  the  neurofuzzy rules  because of
the   overfitting drawback  above. So  we  do  not  have   shining results,
but  the out of  sample    real time  data  will  have  better chance  to
look  like  to  the out of  sample data  used  during the development   test
phase

Anyway,  one  cannot  compare  GA  and  fuzzy  logic  since  they  do  not
belong to the  same  category.
The  first  one is a  learning  method,  the second  is a   method to encode
the  information and the  decision  to  be  learnt. I would  be  like asking
if GA is  more powerful  than Bollinger  bands  what  would  not  make
sense,  they  only  have   here in common the  fact that  they are used   in
trading, regardless to their  results.

One  of  the  most difficult part to understand   with technical analysis
is  the  part  of  illusion  that is in it. GA  will  improve this  part
where  neurofuzzy  is  less  prone  to  that (because   fuzzy  logic
implies  a  logic in  it  through the  decision tree,  what is  less obvious
with the usual  Neural nets  evolved  by  GA.

The  perfect  solution  will never exist anyway. There  is  no  method
leading to  that.
One  may  be  happy  with an approximation  that  understand  most  of  the
obvious trends  and  do  not lose  too  much  money on   noisy  data. It
could  seem  simple  to understand,  but it's  one  of  the   most difficult
task to  make  it for real.  

Neurofuzzy  systems   have absolutely NO  predicting  value. They  just
attempt  to   follow the  trend as  close as  possible in any  circumstance,
what is  sufficient  to make  money ( to  figure it  out, think  to a
centered  moving average,  but  updated to  the  last  bar).
You  may  see  them   here  as  a  kind  of  sophisticated  denoising  tool
applied  to  the market  data serie while  maintaining the  minimum  lag (
by  comparison,  the centered  moving average has no  lag and is a perfect
trading system unfortunately for  past  data only).

Because there  is  no  other neurofuzzy  trading  program  over  there
(unlike  GA  stuff,  widely  spread in the US),  you  will  have some
difficulties  to   get  some valid  feedback information...

Anyway  you  could  check  soon  the real time results  that we  will soon
post on the web site.
If they are  good,  there  is  probably  something  valid  behind all  of
this.

And  if  I was sure  was  not the case,  do  you think that I would  take
the  risk  to  publish  such  results after 20  years in this  business ?

Sincerely,
 
Pierre Orphelin
www.sirtrade.com

Disclaimer:  No affiliation with any GA  program since  1994.
  

 
-----Message d'origine-----
De : David Pyle [mailto:dpevergreen@xxxxxxxxx] 
Envoyé : lundi 24 août 2009 20:49
À : omega-list@xxxxxxxxxx
Objet : [english 100%] Genetic Programs and Neurofuzzy Logic

Dear List, 

I would like to ask what I hope are fair questions regarding genetic
programs and neurofuzzy systems. If I am too simplistic in my understanding,
I hope you will take the time to help me understand.

Here goes.. Are these types of programs super optimizated systems or
distinct systems? Aren't these just brute optimizations, how do they have an
advantage going forward? Don't genetic programs and neurofuzzy systems have
a danger of curve fitting also? In other words if they don't have predictive
features then aren't they just really good at curve fitting?

Thanks for allowing me to ask honest questions?

Dave Pyle