|
PureBytes Links
Trading Reference Links
|
Dans un courrier daté du 05/11/98 21:07:10 Heure d9iver Pari10 Madrid,
jrt@xxxxxxxxxxxxxxxx a écrit :
>
> I am not so much interested in hearing MB's comments as they tend to
> be biased.
>
Woah...
To the credit of MB , he could be sometimes unbiased...
> I am rather interested to hear from Pierre how it is possible to
> distinguish valid from invalid indicators with his software. I have
> asked this question before.
>
Yes I remember.
The answer is difficut to give because we use a combibation of a lot of rues
that depends of all the indicators.
If and indicator has a poor predictive value (undrerstand that we always speak
of the indicator used in conjonction with others), it will produce rules that
are mostly inactive of in conflict.
The current version provides a detailed analysis of thes values, rule by rule,
and the answer can be extracted from this wide array.
This is for the systematic answer.
An other method (qualitative) is to observe the distribution of fuzzy sets
with this version.
They are graphically displayed, so the more fuzzy sets you have for an input ,
the more it produces ( contributes) to the fuzzy rule base. This is a
symptomatic information that let you think that this indicator was really
useful ( fuzzy sets are added in this new version where they are necesary and
IF they are necessary).
The best method is to stick with the general philosophy of the software that
is to code the rules for you from the indicator set, provided that you use
indicators that make sense to YOU when you decide to use them as inputs to the
system.
Even if neurofuzzy logic do not make sense to most of you ( this AI technique
is difficult to use with success without the Assistant) it behaves exactly how
you behave when faced to the blank page of the Power Editor.
The first step in both cases is to have a battle plan, and this is where you
determine wit a correct rpobability if the indicator will be useful.
If you rely on the software to choose the correct indicator by trial and
error, you will waste your time .
> > house. Just throw a few of your favorite indicators in.......close
> > the top and let your new fuzzy friend run on spin cycle......and
>
Your favorite indicators may make sense to you when you decide to combine them
or let the software combine them.
In any case, success will not be met if you do not respect these guidelines.
> The main danger for anyone using such a fuzzy engine is throwing
> heaps of variables at it and it will come up with a solution.
>
This could be right, but we have limited the search space to the training set
dimension.
You will not have one different rule per bar ( perfect overfitting).
If it was the case, generalization will not occur, and you will see this by
yourself when testing on unseen data.
> I have no doubt that Pierre's software will come up with valid
> solutions if given valid predictors. I am sure that a fuzzy engine
> will combine them correctly.
>
> However given invalid indicators it will come up with valid results
> as well. This is why such a tool doesn't make life necessarily easier
> if you don't know what your input is.
>
Right. Therefore the importance of the battle plan.
Invalid indicators could find a local solution by chance, but they will not
generalize on unseen data
To avoid this pitfall, please see the guidelines below.
> Therefore I would be interested to hear from Pierre which techniques
> he uses for prefiltering his input variables.
>
We will translate prefiltering by battle plan.
Here are some example:
A) First plan: TRENDS approach
=========
The market can be seen as a long term short term trend interlacing.
We will use to apply it:
1) A long term trend indicator:
DmiPlus( 25) - Dmiminus(25)
2) A short term one
DmiPlus( 7) - Dmiminus(7)
3) A first difference of one of above ( to tell to the neurofuzzy software
what is the speed)
4) A Trend Quality indicator
ADX(14)
5) It's first difference ( to tell it the trend is established or reversing)
Lookback periods are given as example.
We also check that the theoretical moves to be learned ( zig zag indicator)
are consistent with the indicators moves :
Display the indicators on the chart and ceck by eyeballing
Alternate Plans (A):
You may consider Dmiplus and Dmiminus as separate intputs.
You may replace them by other oscillators ( respect the long / short term
information).
You may consider a Long Medium Short term approach.
Try to saty with less thant 7-9 indicators to avoid unnecessary trainig time
and overfitting.
B) Second Plan:
TREND / TRADING appproach.
======================
You need to tell t te software that your intension is to separate trend mode
and trading mode.
You will use:
1)A trend / Trading indicator:
For example ADX(14)
2) Its first difference ( same idea than above)
3) An oscillator for the trading range
For example RSI(c,5)
4) Its first difference.
5)A tend direction indicator ( DmiPlus( 10) - Dmiminus(10)
Do not forget that the length of the indicators must be consistent with
temoves appearing on the zig zag.
It's stupid to make learn a lot of small trend of 5 bars on a verage with a
DmiPlus( 25) - Dmiminus(25).
C) Third Plan
Cycle mode / Trend mode.
===================
You will need a cycle analysis software (Mesa, MEM, Lomb periodogram).
1)Determine the main period
2) Use a bandpass indicator or one the methods described by John Ehlers
(TAS&C magazine) that will provide cycle signals.
3) Use a trend mode indicator based on the main cycle period.
4) Use the Power spectrum as a trend mode filter, or the Lomb periodogram
probabilityoutput as possible confirmation.
4)Use a trend direction indicator.
The idea is to let the neurofuzzy software to pick what is the trend mode and
the cycle mode and apply the proper signal.
All these tools are available from our web. MEM is also sold by Ehlers ( Mesa)
, Bill Brower, Walter Bressert.
D) Fourth plan:
Market Profile approach.
===================
Use the MP statistics from the MP User function ( we sell this too, we are
worse than an american pharmacy in this domain) to determine the trend trading
mode.
Sveral indicators are useful :
Trade facilitation factor , % of tick above / below distribution , Rotation
Index, first difference of some of them.
Once the trend indicator and trend/ tading mode established with some of the
above, aplly a trend indicator from plan A or plan B and a trading indicator
(oscillator, cycle method fro plan A or plan B).
The Market Profile user function is a part of our MP package.
The other solution are MP (TPO 5) from Mark Brown / Rich Estrem, but I do not
think that they have this kind of function avalable.
Tickscape could be also a solution if it can export to Easy Language.
E) Fifth plan
Equity Curve trading approach
======================
Use a trend system result as a first input (+1= buy -1= sell). Needs only 2
fuzzy sets for this bolean variable.
A moving average crossover, for example
Use a trading range system as second input. (same as above)
( Any OB OS sytem will do the trick)
Use the equity curve position ( from its moving average, for example) as a
measurement of the trend trading range curent efficiency
This approach need a special DLL that will be available soon /or directly with
TS, do not know yet.
Can be replaced by some global variable stuff and several windows running the
2 basic systems.
( see Doug Deming / Kim Horn products)
F) Sixth Plan: Mark Jurik's Indicators special.
================================
You may consider the following idea:
Use CFB as a replacement of ADX
Use VEL (2 differents) as a replacement of long term short term indicators.
Do not forget some first diferences of these indicators.
I could describe dozens of ideas or variant of the above.
I have kept here the wide picture of what technical analysis based on
indicators could be.
IMPORTANT NOTICE:
Choose the training period with uptrend , downtrend and sideways periods.
TO SUMMARY:
+++++++++++
You have seen that in any case I had a precise idea of what interaction could
exist when etablihing a battle plan.
The only anvantage of the neurofuzzy software is that it will write hundred of
rules in seconds, optimize them and produce a very precise interpretation of
the indicators.
It's a time saver that is never tire, and it was because I was tired to do
handcoding and backtesting that we have developped this one.
No magic at all in this.
It's really an assistant and not a PhD.
But it will do in acouple of hour the work that you hardly imagine to do in
one month without sleeping ( backtesing included on unseen data).
As yo may see, I have no secrets to hide here.
All of above is available or feasible for free with some exception ( i need to
upload the DLL's trial to the web in a few days).
I'm a scientific person and not a vendor first.
My choice of sharing most of my knowledge is a definitive one.
Despite what some may think about me, I keep my promises and still wait the
proof of the contrary.
I'm not interseted in money first.
I'm not interested to become a recognized guru.
I'm interested to be paid if my service can help you to make money.
If not, please forget me.
Remember that I do not need to sell all of this and I can make a living with
what I already have.
One Battle plan similar to A or B is explained in details in the web page:
www.sitrade.com/fuzproof.htm
Hope this helps.
Sincerely,
-Pierre Orphelin
www.sirtrtade.com
|