Data reduction in backtesting...

[holyLemon]

... was wondering if anyone had any tips regarding this?

I need to test quite a few stocks starting fro 8 years ago and was hoping to save some calculation cycles by using data reduction and was wondering which techniques are typically used in finance?

Regards
HL

 

[tech/a]

Re: Data reduction in backtesting ...

8 yrs of data isn't normally a great deal in one bourse.
Unless its tick data??
Is it the quantity of data you are concerned about or repetition?

 

[holyLemon]

Re: Data reduction in backtesting ...

hi tech/a,

sorry for my late reply!

Yes the quantity of data i'm concerned about, as one iteration of strategy-evaluation will be surrounded with several loops for determining trading rules in an adaptive way.

I have found some answers to this issue in the meantime however, can do some random sampling (if necessary).

Right now also considering how to construct the system. Thinking of using Matlab as a central controller, with it shooting off tasks to Java & C++ modules (some libraries I have in Java, others in C++). But it may slow things down will need to investigate it.

Do you make use of mathematical packages like Matlab if I can ask?

HL

 

[tech/a]

Re: Data reduction in backtesting ...

hi tech/a,

sorry for my late reply!

Yes the quantity of data i'm concerned about, as one iteration of strategy-evaluation will be surrounded with several loops for determining trading rules in an adaptive way.

I have found some answers to this issue in the meantime however, can do some random sampling (if necessary).

Right now also considering how to construct the system. Thinking of using Matlab as a central controller, with it shooting off tasks to Java & C++ modules (some libraries I have in Java, others in C++). But it may slow things down will need to investigate it.

Do you make use of mathematical packages like Matlab if I can ask?

HL

I dont but Kris my son who is a Doc in Physics with an interest in system design and testing has mentioned Matlab.
I have been using Metastock and Tradesim.
For 12 yrs and am happy with my current systems.
Im a builder not a whizz.
Kris's view is Amibroker Tradesim at this point.
Weve not done much yet as time for both is the issue.
Niether of us need the resultant systems---more to life than trading.

 

[asxiq]

my 2 cents ->
1)rather than fixated by the number of years be it 8 or 4 or any other number ... look for the number of trades the system generated and arrive at statistical accuracy figure which you are comfortable with.

statistical accuracy is 1 - 1/SQRT(n) , n being the number of trades generated by the system in this case ..

so working backwards if the system generated 25 trades you are working with 80% statistical accuracy , for 100 trades : 90 % accuracy , 1000 : ~3% , 2000 :~2%

2) instead of dirtying the hands with matlab ( if trying for the first time to build and test trading strategies) , look for R-package , an open source and free package and the community is great out there