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

Re: Re: Fuzzy Logic vs. System Dynamics



PureBytes Links

Trading Reference Links


----- Message d'origine ----- 
De : Bilo Selhi <citadel@xxxxxxxxxxxx>
À : pierre.orphelin <pierre.orphelin@xxxxxxxxxxxxxx>; <omega-list@xxxxxxxxxx>
Envoyé : samedi 19 août 2000 02:31
Objet : Re: Re: Fuzzy Logic vs. System Dynamics


> ai / nn trading toolboxes galore here they come:
 > 
> there is also a magazine:
> http://ourworld.compuserve.com/homepages/ftpub/jcif.htm
> 

Unfortunately discontinued ( Journal of Computational Intelligence in Finance formerly Neurove$t Journal). 
I have the complete collection, and they still have  back issues I suppose.
Very methematically oriented.
Mark Jurik  has been a first time contributor.

> also check this out for NFL links that has it all:
> http://elve.le.ttu.ee/parveto_www_home/MESEL/R&D/NEUROFUZ/RESSOURC.HTM
> 
> Pierre,
> i am still evaluating your package... so i can not comment
> on it.
> 

Plase do...


> *my favorite is this tool:
> www.sdi-inc.com  link is out they might have moved or down
> if someone could let me know if they moved.
> anyway, it's ga based app called E.
> you give it a set of price data and a set of inputs ( price derivatives or
> indicators ) and you let it run. it trains based on max profit and
> when it's done and found the best solution it outputs:
> - variables
> - functions
> - rules!
> now, that's something, especially rules...
> now if you explore it long enough you can find some rules that could
> point you to a theoretical solution...
> also you can code those rules right into TS. ie the program finds the
> best rules, variables and functions for you...
> the tool is still in it's infancy but - what a tool and what a concept!!!
> smoking... this is what research should be...
> they call it "stock trader scoring" ie. each hypothetical trader is evolved
> based on the amount of money "he" accumulates. when the best trader
> ( solution ) is found - it spits out the rules by which the most money was
> made...
> it's used now in chemical engineering, electronic circuit design, etc.
> f...king thing can generate a complex circuit based on desired input and
> output
> without human interference just by using a set of available basic circuits
> ( set of available functions ). what a concept... what a tool...
> bilo.
 
In 1994 we have developped such a tool ( Galvano) for TS.
Was able to code plain easy language rules from a powerful GA.
See an example below.

5000 g,n,rations(/g), 2350 chromosomes(/c). 1 famille(/f)
³Hybr(/h):29.9% Mut(/m):19.4% Inv.Pt(/p):19.4% Inv.Ch(/s):9.7% Lamarck(/l):1.2%
³Eug,nisme(/e):N  Intensit, s,l.(/i):6.0  Hybr XOR Ops(/o)  Equil. auto(/q)
³CopieS,l.(/0):O Op.S,l.(/1):N Hybr.S,l.(/2):O TxOps(/3):80% HybMini(/4):30%
ÀTemps maxi(/t):0s Info toutes les 20s(/d). 1 chr=66 octets. 413kO,reste 640
GABRF-15 rSgles(/x) (35b). Formulation(/y):3. 
Classe voulue(/w):"S-1", 7 entr,es, Mode rapide(/+), base(/=)=0.50,
zoom(/==)=1.00. Noter sur taux de recl(/$0). Equilibrage(/*). D,tailler les
mouvements dans le fichier(/DO). D,tailler les mouvements dans le systSme
obtenu(/DR). 

time = (time - 905.0) / 695.0 ;
fractal1 = fractal1 / 80.50130 ;
fractal2 = (fractal2 - 19.47360) / 49.07380 ;
adx = (adx - 9.74940) / 51.49630 ;
Dmi+ = (Dmi+ - 3.00370) / 65.97780 ;
dmiminus = (dmiminus - 2.08490) / 70.01820 ;
volume = (volume - -87.07890) / 165.45780 ;

IF  (fractal1 - 0.12) < (adx + volume) AND
 ( (0.56 + 0.56) < (0.56 * 0.64) OR
   (adx - 0.10) < (adx + time)) AND
 (0.73 + 0.02) < (fractal2 + 0.49) AND
 (Dmi+ * 0.37) < (0.17 * 0.76) OR
 ( ( (adx - time) > (dmiminus - 0.00) OR
     (0.27 - 0.56) > (0.20 + 0.59)))

THEN S-1

ELSE S1

Strange rules, I know, but the thing was able to do more ( if cos( dmi)+7.3654 >squareroot(volatility) AND....OR 
Very funny and was perfectly working on training data.
Generalization was not so  good than with NFL  but quite acceptable.

The same program had a  GA scorecard and a multidimentional no layer NN ( all cells connected in 3D, like in a true brain).
All of this was running in TS3, TS4, all of this trained by GA, and producing trading systems.
The GA had the latest  features ( evolved several families in parallel, exotics GA operator, and very fast - was a DOS program)

The programmer was probably a genius, but very difficult to live , so I gave up in 1994 and started to work with JewelSoft.

Here is a scorecard system by GA
(Sorry this version was speaking french, but there was also an english version, but I have lost the  results files).
No guarantee for ths system provided here.
I do not remember of the results. It's only past history for me.

3040 chromosomes(/c), 50000 g,n,rations maxi(/g). 1 famille(/f)
ÀTemps maxi(/t):0.0 secondes Information toutes les 10.0 s(/d) Stop . 120:Non
ÚEquilibrer les taux(/q):Oui, les sites(/n):Non.  /NonZero:Oui, 3 essais
³Ajuster le taux d'hybr.(/qh):Oui. Taux maxi permis(/OLim):0.40.
³Ajuster le taux total(/OVar):mode On stats, G1.5(/QG).
³Elitisme(/Eli):Oui  Eug,nisme(/Eug):Non  Intensit, de la s,lection(/i):2.5
³Biaiser: Copie(/BsC):Oui  Hybridation(/BsH):Oui  Autres Op,rateurs(/BsO):Oui
ÀTaux d'Op,rations(/OMin,/OMax):20-90%  Hybr.mini(/HMin):15%  Hybr XOR Ops(/o).
M,moire utilis,e: 534kO, reste 640 disponibles. 1 chromosome=70 octets. 

Galvano Score - Pond,ration Maxi=171, 6 plages(/y) (variable, /z).  Seuil
mobile(/#). 
Classe voulue(/w):"S-1", 7 entr,es, Mode rapide(/+), Taux ou gain minimum
admis(/mini)=0.50, facteur de zoom des justesses(/zoom)=1.0.
Noter sur taux de reclassement(/$0). Equilibrage de S-1 et S1(/*). D,p"t:1.000
(/dep). 
Fitness limite: 128(best_perf=0.82)   (gic_limite_reclass=0.76)
NbOK Voulu Appr: 886QF+1475NF N708/937, Q302/400  Test: 0QF+0NF N79/105,
Q79/105
Justesses:Mini=0 Modale=153 (153) Moy=124.3 Maxi=158 (G6, Sat=0.61).
Taux: Hybr(/H):9.4% Mut(/M):17.6% Inv.Pt.(/P):21.4% Inv.Ch.(/S):6.0%
Lam.(/L):6.1% Simplex(/Bers):2.4%  - Total:63.0
Tapez "GASCF ?" sous DOS pour des d,tails sur toutes les options.




Grille de score nø 1
Seuil: 654.0

time        Bornes: 905.0   1010.3    1600.0    
            Points:    18.8      112.0     
fractal1    Bornes: 0.00000 24.26066  41.90478  59.54890  80.50130  
            Points:    85.2      58.3      2.7       103.9     
fractal2    Bornes: 19.47360 32.55995  68.54740  
            Points:    37.6      126.1     
adx         Bornes: 9.74940 17.58579  33.25858  42.21446  46.13266  61.24570  
            Points:    152.2     125.4     106.6     134.8     171.0     
Dmi+        Bornes: 3.00370 4.16121   24.99630  36.57135  68.98150  
            Points:    142.8     125.4     2.0       15.4      
dmiminus    Bornes: 2.08490 24.36342  25.95474  29.13739  59.37252  72.10310  
            Points:    75.8      77.8      60.4      105.3     113.3     
volume      Bornes: -87.07890 15.01208  18.53246  78.37890  
            Points:    117.4     140.2     44.3      

S,lection des "S-1".


----------R,sultats du test sur gbrdemo.lnf: 81.6 %

Pr,dictions correctes et incorrectes:
Nombre de "S-1" bien pr,dits: 234
Nombre de "S1" bien pr,dits: 520
Nombre de "S-1" mal pr,dits: 43
Nombre de "S1" mal pr,dits: 140


Sincerely,

-Pierre Orphelin
 web: http://www.sirtrade.com