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

Re: Gambling Indicators: They work!



PureBytes Links

Trading Reference Links

Dans un courrier daté du 08/10/98 19:11:33  ,  Gabe Hanover écrit :

<< Your comment:
 
>>>The core problem for traders is that they believe some indicators to be
 better than others. This is not true because an indicator is [only] a
 measurement of some [of the] specific characteristics of the underlying data
series
 (raw data). The misleading point is in the way [we] use them. [M]ost of us
fail
 because the rules given with the indicators are too simple to encompass the
 market complexity.
 
 is the truest statement I have yet seen regarding market systems. I am
 going to put it on the wall to remind me to not look for the holy grail
 indicator, but keep working on the overall system concept. That is,
 first break up the complexity of the market into its many identifiable
 phases (for exapmle, expansion-contraction, trending-choppy, etc.). Then
 find the simplest (least curve-fit) entry-exit indicator which will
 properly handle each phase. Finally, determine the optimal position size
 for each entry.
>>

Breaking the problem into smaller sub problem is THE solution when using
indicators.
Unfortunately, we displace the problem into others difficult to resolve by
hand coding:

-How  much break points  for a given indicator ?

-Where to place the boundaries ?

-How to build the rules that can be numerous (in fact N= product of  the
number of sets for each indicator).
Means that if we break 3 indicators into 5 sets for each , the number of rules
is 5^3= 125 rules.

-How to deal with the at boundaries problem (the classical RSI 30 /70
overbought oversold is the perfect trivial example).

-Where to add new sets do deal with the  intrinsic indicator meaning ?

 for example  a RSI will maybe not  need a 5 parts breaking sheme like this:

0|__a______|___b_____|___c_____|__d______|__e______|100

but rather
0|____a_______|__b__|__c__|__d_|________e__________|100

or
0|____a____________|_b_|_c_|__d_|_______e__________|100

There is an infinite number of solutions...

Now , the same problem will exist with the others indicators that you will
have to combine wt his one.
You could spend your life to write and test such systems (that are really the
best way 
to interpret raw indicators).
We have recently solved  all of the above points with the new version.

NNets fails here because they do not deal with the set concepts and boundaries
problem (no problem with boundaries with N inputs , because they are treated
continuously througn the activaton function ( sigmoid and the like).

Suffice to read the last Omega Magazine Issue ( Murray Ruggiero interview)
wherehe clearly said that he gave back with NN and indicators and pleaded for
an indirect use.
And he is not alone in his case.
I found the same  deceptive answer years ago, by using all the TS compatible
available  NN products.

 What is amazing with the survival NN products (indicators cruunching based)
for trading is that they try to survive by adapting:

Neuroshell Trader (Wards systems) has a similar training schem than ours (
trains on a performance summary field like Net profit instead of the old
uncorrelated RMSE), but their ads claims: We have  a customer that made
millions $ withour products, but it took month to find the good indicators!
If the solution was so obvious with NN and indicaors , you should find out in
a few days...
That's aside, Wards NN products are very serious and of a good quality ( I
used Neuroshell 2 and helped Steve Ward to develop the first Neuroshell2 TS
interface).

You may check their web site<A HREF="http://www.wardsystems.com";>
http://www.wardsystems.com</A>

The other side is Biocomp Profit.
They use Genetic algorithm to train the NN.
We have made this years ago, but it's not so easy to get it to generalize.
They provide a free evaluation version, I believe, but unfortunately, the ad
they publish shows than  on unseen data, it finds good results.
Examining the ad text (or the web site, I do not remember) , you will see that
they use synthetic data , and the test part is generated by the same method
(trend + cycles + noise added). So it's not a real world example!
That's aside again, Biocomp products have also good reputation. But they are
not traders.
See  http://www.biocompsystems.com

The problem is that developping NN and trading systems is not the very same
business, and there are few people abe to do this (Bob Brickey is one of the
few, in his own domain - I know that he does not like indicators! -).

I do not know very well AI software from a programming point of view. Maybe a
little bit more than the average AI skilled trader, but not much more.

I cannot say the same from  the indicator knowlegde that I  always consider
from a physical point of view ( a physical problem solves from masurements
done, provided that they are meaningful) .

This is the reason why I work with AI specialists in their domain (neurofuzzy
logic in our case). The result is what you have seen on the web page.
It's a team work where  both parts should have been unable to produce
something interesting without the help and knowledge of each other.

<<
 want to thank you for shining a light on the path we should be taking
 and reminding us we face a task no less daunting than NOAA has in
 predicting the weather.
>>
Thanks,but it's easier to build trading systems with such techniques than
predicting the weather at this time.
But what 'I'm saying here ? My God, I forgot that I was not believing... 

Sincerely,

-Pierre Orphelin
The explanation are still  available on 
<A HREF="http://www.sirtrade.com/fuzproof.htm";>
http://www.sirtrade.com/fuzproof.htm</A> for those who missed the thread.