Mikkom,
After reading whole thread twice, i think i am quite familiar with (in general) what you are doing. It's time to ask more specific questions
How I see it:
1. Generate huge pool of strategies, based on events, from limited space. Discard those ones, which doesn't meet certain (simple?) criteria.
2. Find few strategies from pool, which combined together, has best fitness (which doesn't mean highest profit).
Ad Events
- How do you code events? Just occured/not occured (1/0)?
- Are events related to price movement only? Do you incorporate daily min/max distribution (presented in the beginning of the thread) in some manner?
- I assume that each type of event has many "mutations" (combination of some parameters). How many approximately? How many parameters? Is this what do you mean by limited space - Few type of events (2+2?) (defined by you), few type of parameters (defined by you), set of values for each parameter (defined by you)?
- How long does it take to create event file for one instrument? (I know, it isn't important since you do it only once, but I am curious
- Does event-based trading implies fixed SL, no PT?
Ad GP(general)
- How big is your population?
- How many generations?
- Do you use ADF? (automatically defined functions - term from Koza's books)
Ad GP(stage 1)
- Terminal set is obvious (events) but What functions are in function set?And, or, if, not?
- From snapshots provided I think that creating pool of strategies has incremental nature. Do strategies already in pool have any influence on candidate strategy's fitness? I.e. discard strategy with high average correlation with strategies in pool? Can such a method contribute to pool diversity (is diversity desirable?) and prevent DD clustering?
- Is the "top X trades generate max Y profit" rule currently present in the fitness algo?
- Is strategy profitable only on one instrument valid?
Ad GP(stage 2)
- Do you weigh each strategy, or you take it from pool "as is"?
- What is in a function set? I can't imagine other function than "plus" (or "plus" and "times" in case of weighing)
Ad in sample/out of sample (IS/OOS)
- IS data must be large, and fitness algo must be constructed that way, that high fitness in IS leads (usually) to high fitness in OOS. When discarding strategies with high fitness in IS and low fitness in OOS, we are in fact fitting also to OOS (is this case of "survivorship bias"?). Do I interpret your words correctly?
Ad whole system
- How long is your average trade? How the distribution looks like?
- How many trades are you simultaneously in on average?
- What plays greater role in equity smoothness? Diversity of strategies or diversity of instruments?
Huh, 31 questions in one post Any answer (especially those regarding GP) will be highly appreciable
Thanks
After reading whole thread twice, i think i am quite familiar with (in general) what you are doing. It's time to ask more specific questions
How I see it:
1. Generate huge pool of strategies, based on events, from limited space. Discard those ones, which doesn't meet certain (simple?) criteria.
2. Find few strategies from pool, which combined together, has best fitness (which doesn't mean highest profit).
Ad Events
- How do you code events? Just occured/not occured (1/0)?
- Are events related to price movement only? Do you incorporate daily min/max distribution (presented in the beginning of the thread) in some manner?
- I assume that each type of event has many "mutations" (combination of some parameters). How many approximately? How many parameters? Is this what do you mean by limited space - Few type of events (2+2?) (defined by you), few type of parameters (defined by you), set of values for each parameter (defined by you)?
- How long does it take to create event file for one instrument? (I know, it isn't important since you do it only once, but I am curious
- Does event-based trading implies fixed SL, no PT?
Ad GP(general)
- How big is your population?
- How many generations?
- Do you use ADF? (automatically defined functions - term from Koza's books)
Ad GP(stage 1)
- Terminal set is obvious (events) but What functions are in function set?And, or, if, not?
- From snapshots provided I think that creating pool of strategies has incremental nature. Do strategies already in pool have any influence on candidate strategy's fitness? I.e. discard strategy with high average correlation with strategies in pool? Can such a method contribute to pool diversity (is diversity desirable?) and prevent DD clustering?
- Is the "top X trades generate max Y profit" rule currently present in the fitness algo?
- Is strategy profitable only on one instrument valid?
Ad GP(stage 2)
- Do you weigh each strategy, or you take it from pool "as is"?
- What is in a function set? I can't imagine other function than "plus" (or "plus" and "times" in case of weighing)
Ad in sample/out of sample (IS/OOS)
- IS data must be large, and fitness algo must be constructed that way, that high fitness in IS leads (usually) to high fitness in OOS. When discarding strategies with high fitness in IS and low fitness in OOS, we are in fact fitting also to OOS (is this case of "survivorship bias"?). Do I interpret your words correctly?
Ad whole system
- How long is your average trade? How the distribution looks like?
- How many trades are you simultaneously in on average?
- What plays greater role in equity smoothness? Diversity of strategies or diversity of instruments?
Huh, 31 questions in one post Any answer (especially those regarding GP) will be highly appreciable
Thanks