|
PureBytes Links
Trading Reference Links
|
Dans un courrier daté du 22/07/98 18:00:33 , TC écrit :
<<
Orphelin@xxxxxxx wrote:
> A track record is the common proof of the ability of the supposed
successful trader.
> >From a R&D point of view, it's a nonsense.
I think more so the track record shows the proof of two things :1. A good
system
2. The ability of the Captain to steer the Ship!
>>
A track record is a costly (time) and unsure proof.
If someone wants to sell a system that has a good track record, and a good
bactesting record, he will do.
This is only the proof that the system has been used, not of its robustness.
You may have an overptimised system on the backtest period and a good trading
part of profits during the track record period.
The results can be good also by chance and do not show the overoptimization
period .
<<
> IF a system is correctly bactested AND on sufficient amount of data that
makes
> chance or randomness nothing as a part of the process, the need of a track
> record is worthless.
How many Pierre? can do this! They all want to win so bad that they taint the
process with the
desire to do so. This is a point where all trading knowledge will in fact
keep them further from the
truth. Honesty will lead the way if they will let it. "The Truth is Out
There"
>>
Depends partially on the number of degrees of freedom of the system.
When developping with GA years ago, we have found that the 20,000 or 30,000
bars range were very difficult to optimize, and still able to perform on
unseen data, regardless of the system developped.
GA was used to build scorecards, rules bases systems, and n dimentional NNets.
Same kind of results regardless the underlying structure. Too strange not to
be appealing...
So, in my opinion, there is also a limit to the degrees of freedom influence
when comparing training and test data, because the market, even if very
complicated has some basic dimension tht cannot be encompassed with too much
degrees of freedom.
The key is huge amounts of data.
So, there is also a limit to overoptimisation when you get a system working
also on a serious amount of unseen data.
I think that I I had not made these extensive tests years ago with GA
software, I would not have "discovered " this, that seems not widely known
(and that definitively rejects the need of a track record, as I'm able to test
over 50,000 unseen bars if I want)
<<
> Means that you need testing on 20,000 bars of unseen data or more.
Please make an exception for me I only have 17,500 points of 7 hr. bond data,
prior to that it was a
6 hr. day. But I agree none the less! I no longer laugh at the system results
published here with 6
weeks of data. Now I only snicker, it has lost its effectiveness on my humor.
It will in time, I'm
sure become saddening for me.
>>
6 weeks of data for testing is a joke, as well as for a system generating
hundred of trades over a longer period. We need thousands AND thousands of
trades on unseen data.
Now, to partially overcome the lack of avialable historical data, you may
either test over other markets, or if the sytem is specific, on data that have
a different compression rate.
For example 60 min data and 50 min data and 70 min too. Or 68...
The waveshapes are nor strictly the same, and if the trick does not really
multiply by a factor 3 the size of the database, its an acceptable workaround,
better that staying only with the original 60 min bars.
<<
I know you were and very few others. I have learned some things from you and
I appreciate you.
>>
Same for me.
We will stay outside of the (too sensible) Omega Reserach topic.
<<
About bad experiences with commercial BS codes:
>This is how you got your black color, and you forgot this.
>>
Your correct I have forgotten, My intolerance grows with time, I should start
a foundation for
market study so the hard edge may be smoothed out some. I have been thinking
about this for some
time. In fact I have been thinking about a Unix based development platform
for trading system
development, I have a server commitment already, just need the analytical
half of the program. I
would make it available to Universities and qualified individuals.
>>
Interesting project.
Intolerance has nothing to do with this.
No need to be intorerant, if you have got the knowledge, regardless to the
price paid.
This is where real superiority is.
<<
About security and publishing free systems for trial:
If I were certain of security, I would consider the same. The mere fact that
you don't publish now,
says to me you have something good else you wouldn't worry about security. I
in fact hope to be the
recipient of your demo's, I am now more than ever money motivated and not
trader motivated, So if
one of your systems were to smooth out my cumulative equity curve, I would of
course purchase it.
>>
I'll see what happens with TS5 protection schemes and DLL's.
For the moment it's absolutely not secure (TS4).
Outside of this , I do not consider to propose something that one cannot test
by itself
I have some(a few!)systems performing well in my opinion (and I coded several
thousands of pure BS of my own).
Sincerely,
Pierre Orphelin
www.sirtrade.com
|