Designing a System from scratch

[beginner_av]

Timeframe: I would be starting with EoD data, but may switch to whatever timeframe gives me a trade that will last for average 2-3 days, and max 5 days, I look at this as a requirement and not as an assumtion.

Simple

Trading EOD is fine, but system development IMHO is not fine. Use tick data or at least 5 mins data sos sys dev. For big range days, you will never know what (H or L) came first and you will be taking anywhere between Open at High to Open at Low and Close at High to Close at Low. The CAR (even if all other stats look acceptable), can vary by way over 100% and more by using this blind data.

 

[beginner_av]

CV do u read this anymore. Why dont you give a thread on costs of systems development? (Will scare half the people away!)

 

[beginner_av]

Still not sure what to post next lol.Perhaps a round up on the various Performance Measure used in analyzing trading systems would be a good idea.

Also perhaps a few words on a 'premise' focussed trading system development vs a 'if it works it must be good' system development.

Anyways, will be back soon

Rgds
CV

May I add a list Sir CV?

a. Data division

shhh!!! (cant give too many other things away. will be strictly between us! )
[for non CV readers - pure pun intended!]

 

[oxusmorouz]

Hello CV,
Can you please comment on the "utility" testing method I posted here:
http://www.traderji.com/111437-post26.html

Thanks

 

[beginner_av]

Very good question. Simple answer is that you have to looka at it from system performance parameters and adjust your capital allocation accordingly. Let CV give a more precise answer.

 

[oilman5]

an excellent thread to read on.
regards to all

 

[kkseal]

What would be the MAIN parameters for judging system performance? In other words what sort of results, based on which parameters would qualify a system as good or bad?

Regards,
Kalyan.

P.S. : I only have Ami & EOD data to go by.

 

[kkseal]

Another question ...

What would be the right structure of a coded system?

CV mentioned modules. Well they do have their place; but the broader structure i envisage (& one emerging as a natural outcome of my thought & development processes) is that of a multi-level decision tree with multiple nodes at each level. The decision ('yes-no') at each level will be based on one or more Technical/Fundamental parameter (preferably one main & 2 supporting/confirmatory parameters) implemented through 'plug-in' modules (this will enable me to shift/re-arrange the nodes more easily, if needed).

[This decision tree can hopefully then act as the 'feedstock' for a neural net at a later stage - though i'm not too sure about that due to my paltry knowledge of nu-nets at this stage]

But i'm not satisfied entirely with this structure. What i'd like to achieve is a clean segregation of each level based on the specific function it performs (i have only one such level as of now - the Fundamentals level; the others are kind of 'mixed-up'. While i'm not saying (as of now) that it's unachievable within the current (aforementioned) structure, what's increasingly appearing to be a 'cleaner' design is that of MULTIPLE decision trees each based on a specific function (of the entire decision process) and each with it's own levels & nodes (& sub-levels & sub-nodes, if necessary).

Modules (as mentioned by CV) would be the 'bricks', the building blocks, of such a structure but what would be more important is the 'cementation' - between levels & nodes in the first structure and between decision trees and the levels & nodes within them (not to mention sub-levels & sub-nodes, if needed) in the second structure.

Regards,
Kalyan.

 

[CreditViolet]

Another question ...

What would be the right structure of a coded system?

Before coding, these are the things to keep in mind

http://www.traderji.com/attachment.php?attachmentid=7016&stc=1&d=1194639643

Basically seperate Initial Bet Sizing, Decisin Logic and Trade Management i.e Exits.Should be straightforward for easy debugging.

CV mentioned modules. Well they do have their place; but the broader structure i envisage (& one emerging as a natural outcome of my thought & development processes) is that of a multi-level decision tree with multiple nodes at each level. The decision ('yes-no') at each level will be based on one or more Technical/Fundamental parameter (preferably one main & 2 supporting/confirmatory parameters) implemented through 'plug-in' modules (this will enable me to shift/re-arrange the nodes more easily, if needed).

[This decision tree can hopefully then act as the 'feedstock' for a neural net at a later stage - though i'm not too sure about that due to my paltry knowledge of nu-nets at this stage]

But i'm not satisfied entirely with this structure. What i'd like to achieve is a clean segregation of each level based on the specific function it performs (i have only one such level as of now - the Fundamentals level; the others are kind of 'mixed-up'. While i'm not saying (as of now) that it's unachievable within the current (aforementioned) structure, what's increasingly appearing to be a 'cleaner' design is that of MULTIPLE decision trees each based on a specific function (of the entire decision process) and each with it's own levels & nodes (& sub-levels & sub-nodes, if necessary).

Modules (as mentioned by CV) would be the 'bricks', the building blocks, of such a structure but what would be more important is the 'cementation' - between levels & nodes in the first structure and between decision trees and the levels & nodes within them (not to mention sub-levels & sub-nodes, if needed) in the second structure.

Hmm..Interesting..You should look at State Machines and Decision Trees.Check out @Risk Decision Suite, it has a pretty good decision tree module.

Cheers

 

[CreditViolet]

What would be the MAIN parameters for judging system performance? In other words what sort of results, based on which parameters would qualify a system as good or bad?

Regards,
Kalyan.

P.S. : I only have Ami & EOD data to go by.

Geometric Growth Rates, Entry Edge i.e High Favorable Excursion, Distribution Stats, Dependency Analysis and LR of the EC are a good starting point