ATS Tick based manipulation of the OHLC bars, Genetic Optimization and integrated multiple strategies is the Holly Grail
Combining multiple systems
- Thread starter Arrow
- Start date
Dude... don't take it too personal.
I don't have to agree to everything you post in here and am I not allowed to post something against it? If having my own mind is "stupid" or seemingly "know-it-all" then so be it.
I did take the luxury of testing the original Parrondo's example and played around with it. Still not impressed. Regardless of who's right or wrong (Parrondo has value or not...), I don't care much. I did my research and tests, came out with a conclusion, and I'm going stick to it, rather than wasting time trying to find a trivial usage of it.
The topic of combining multiple systems should start from a practical and qualitative view point.
All financial markets are tied together in some way and will therefore share the same risks, but to different to degrees. Therefore our trading systems should theoretically allocate money to each factor in such a way as to optimize the reward/risk for a certain level of risk or in other words to allocate money so as to meet a risk free rate of return for a given period. This may sound like "old hat", but if I reworded it to say how I am looking for a strategy that has a 95% probability of never having a loosing period for x amount of periods, then it becomes a novel idea for some readers.
In order to quantify these risks with there associated reward, it is necessary to categorize them: micro-shocks,major shocks, thematic (subprime, internet bubble, etc.), economic, time of day, correlated , half-life risk,individual system risk, ), liquidity,etc. The half life risk is the risk we face when the exploited inefficiency goes away. This is the amount we expect to lose once our system no longer works.
any reward/risk that is shared by all markets, can be accounted for and is significant should be acknowledged. For our purposes, we could boil it down to descriptive factors, but just know that it would take away from a true optimization. The prior list is more of an example than rigid factor.
The idea is to maximze optimal f (the general idea) in terms of the reward/risk for all the variables mentioned for all systems together.
Unless anybody has objections to this general framework for optimal system alloation, I will continue after I finish moving trees out of my yard.
All financial markets are tied together in some way and will therefore share the same risks, but to different to degrees. Therefore our trading systems should theoretically allocate money to each factor in such a way as to optimize the reward/risk for a certain level of risk or in other words to allocate money so as to meet a risk free rate of return for a given period. This may sound like "old hat", but if I reworded it to say how I am looking for a strategy that has a 95% probability of never having a loosing period for x amount of periods, then it becomes a novel idea for some readers.
In order to quantify these risks with there associated reward, it is necessary to categorize them: micro-shocks,major shocks, thematic (subprime, internet bubble, etc.), economic, time of day, correlated , half-life risk,individual system risk, ), liquidity,etc. The half life risk is the risk we face when the exploited inefficiency goes away. This is the amount we expect to lose once our system no longer works.
any reward/risk that is shared by all markets, can be accounted for and is significant should be acknowledged. For our purposes, we could boil it down to descriptive factors, but just know that it would take away from a true optimization. The prior list is more of an example than rigid factor.
The idea is to maximze optimal f (the general idea) in terms of the reward/risk for all the variables mentioned for all systems together.
Unless anybody has objections to this general framework for optimal system alloation, I will continue after I finish moving trees out of my yard.
Quote from man:
i stopped believing in anything i hear and read until i
tried it out within my team ...... but usually my guys
come back to me: "all crap" ... but who knows, once in
a while ...
my take on this, just from top of my head and without
looking at it: the systems, even if they do loose over
years must have certain features. they can't be random,
since out of random you can't moneymanage return.
martingale blablabla.
this certain feature must be autocorrelation. (sorry
you probably covered all that ... i did not bother reading
the whole thread ... too weird a day out there ...).
now your system of systems just trades that: the
autocorr of the underlying equity curves. all of a sudden
it makes perfect sense.
Note that this autocorrelation is trivial if its just a function of observeable market. If not.. well then you have discoverd something which can be manufactured into something tradeable. That would be the takeaway. Let's not fall victim to the representation falacy. data != tradeable edge, but the idea behind it does.
i am not sure if you misunderstand me: i am talking
about autocorr of the subsystems equity curve. has
no thing to do with autocorr of markets. no thing.
your previous post seems way too theoretical to me.
all that sounds great, but has limited merit in my eyes.
i am afraid you end up avoiding the real hard work to
dig for another edge by understanding dozens of details
of your existing one.
i dare a little piece of advice: you remind me of myself
some time ago. trying to make it "right" and make it
"convincing" and well spoken. here is the advice: get
rid of this attitude. the sooner the better. but no
offense intended. no desire for conflict on my side.
Quote from man:
i am not sure if you misunderstand me: i am talking
about autocorr of the subsystems equity curve. has
no thing to do with autocorr of markets. no thing.
your previous post seems way too theoretical to me.
all that sounds great, but has limited merit in my eyes.
i am afraid you end up avoiding the real hard work to
dig for another edge by understanding dozens of details
of your existing one.
i dare a little piece of advice: you remind me of myself
some time ago. trying to make it "right" and make it
"convincing" and well spoken. here is the advice: get
rid of this attitude. the sooner the better. but no
offense intended. no desire for conflict on my side.
I understood your autocorr. post correctly.. what I was saying was that if the autocorrelation of the subsystems equity curve is not a function of the markets traded, then you have something, otherwise it's like using indicators to look at price.
The previous post was meant to act as a framework. I like this topic and I wanted to propose a starting point. tada
sometimes communication can be tricky in forums.
I guess I've been asking to many questions... 
Though it isn't the only "complication" I add to my own portfolio management, I've briefly wrote about categorizing systems and base tendencies the systems are using. Hopefully adding my own experience can provide me with more feedback regarding it to help me get more ideas about it...
Starting with the obvious, I have my systems categorized by:
1. Product Type (equities, index futures, forex, interest rate futures, options, commodities like oil, corn and etc.)
2. Region (Asia, US, EUR)
3. Timeframe and data (bar compression like EOD, 1 min., PandF, tick, single bid/ask, market depth)
4. System type (trend-following, counter-trend, scalp, High Freq.)
5. Tendency type
6. etc. etc. etc.
Obviously, all my systems are members of multiple categories. I think we can all agree that there isn't a single measurement that is reliable on all cases and we use what's appropriate depending on the each system and portfolio. One of my approach is to have a grand measure for each sub-category. As an example, I would measure the frequency of the positive outlier for trend-following, and use correlation matrices for product type and region. In another words, I use multiple measurements and run multiple tests for each subcategory (as a minimum) along with tests for the whole portfolio.
Up till now, I think what I do is very much average and the norm for portfolio management. I personally know a few people who takes the same / similar approach as me. It's nothing special or new. The tests and results are very much black and white. Based on the results, you get an outline of how you will be allocating capital to the systems.
Obviously, I have a few categories that are very grey. Grey meaning there's no statistical measure to rate them.
As a simple example, I have a walk-forward optimization logic that automatically runs when a specific condition triggers. In another words, it's an self-adjusting logic. I can have 2 systems, that has no common categories from the list above.
Now, on multiple occasions I observe that these 2 different systems adjusts closely. Based on the observation, I run a test that triggers the opt. logic for both systems when one of the system triggers. I got a very positive result. So I have a "system of a system" that re-adjusts it's parameters in multiple systems triggered by other systems.
If I was able to extract the specific tendeny of why the 2 systems are connected, I would have done so and categorized it. But I don't have a logical explanation and I don't have a statistical measurement to detect these kinds of "relationships". All I can say is, it adds value to my trading.
So I'll end this post with another cliche...
Everything depends and there's no magic formula.
TEST EVERYTHING.
PS. Wooohooo!!! I didn't end it with a question. Or should I end it? :eek: ooops...
Though it isn't the only "complication" I add to my own portfolio management, I've briefly wrote about categorizing systems and base tendencies the systems are using. Hopefully adding my own experience can provide me with more feedback regarding it to help me get more ideas about it...
Starting with the obvious, I have my systems categorized by:
1. Product Type (equities, index futures, forex, interest rate futures, options, commodities like oil, corn and etc.)
2. Region (Asia, US, EUR)
3. Timeframe and data (bar compression like EOD, 1 min., PandF, tick, single bid/ask, market depth)
4. System type (trend-following, counter-trend, scalp, High Freq.)
5. Tendency type
6. etc. etc. etc.
Obviously, all my systems are members of multiple categories. I think we can all agree that there isn't a single measurement that is reliable on all cases and we use what's appropriate depending on the each system and portfolio. One of my approach is to have a grand measure for each sub-category. As an example, I would measure the frequency of the positive outlier for trend-following, and use correlation matrices for product type and region. In another words, I use multiple measurements and run multiple tests for each subcategory (as a minimum) along with tests for the whole portfolio.
Up till now, I think what I do is very much average and the norm for portfolio management. I personally know a few people who takes the same / similar approach as me. It's nothing special or new. The tests and results are very much black and white. Based on the results, you get an outline of how you will be allocating capital to the systems.
Obviously, I have a few categories that are very grey. Grey meaning there's no statistical measure to rate them.
As a simple example, I have a walk-forward optimization logic that automatically runs when a specific condition triggers. In another words, it's an self-adjusting logic. I can have 2 systems, that has no common categories from the list above.
Now, on multiple occasions I observe that these 2 different systems adjusts closely. Based on the observation, I run a test that triggers the opt. logic for both systems when one of the system triggers. I got a very positive result. So I have a "system of a system" that re-adjusts it's parameters in multiple systems triggered by other systems.
If I was able to extract the specific tendeny of why the 2 systems are connected, I would have done so and categorized it. But I don't have a logical explanation and I don't have a statistical measurement to detect these kinds of "relationships". All I can say is, it adds value to my trading.
So I'll end this post with another cliche...
Everything depends and there's no magic formula.
TEST EVERYTHING.
PS. Wooohooo!!! I didn't end it with a question. Or should I end it? :eek: ooops...
Quote from TSGannGalt:
I guess I've been asking to many questions...
Though it isn't the only "complication" I add to my own portfolio management, I've briefly wrote about categorizing systems and base tendencies the systems are using. Hopefully adding my own experience can provide me with more feedback regarding it to help me get more ideas about it...
Starting with the obvious, I have my systems categorized by:
1. Product Type (equities, index futures, forex, interest rate futures, options, commodities like oil, corn and etc.)
2. Region (Asia, US, EUR)
3. Timeframe and data (bar compression like EOD, 1 min., PandF, tick, single bid/ask, market depth)
4. System type (trend-following, counter-trend, scalp, High Freq.)
5. Tendency type
6. etc. etc. etc.
Obviously, all my systems are members of multiple categories. I think we can all agree that there isn't a single measurement that is reliable on all cases and we use what's appropriate depending on the each system and portfolio. One of my approach is to have a grand measure for each sub-category. As an example, I would measure the frequency of the positive outlier for trend-following, and use correlation matrices for product type and region. In another words, I use multiple measurements and run multiple tests for each subcategory (as a minimum) along with tests for the whole portfolio.
Up till now, I think what I do is very much average and the norm for portfolio management. I personally know a few people who takes the same / similar approach as me. It's nothing special or new. The tests and results are very much black and white. Based on the results, you get an outline of how you will be allocating capital to the systems.
Obviously, I have a few categories that are very grey. Grey meaning there's no statistical measure to rate them.
As a simple example, I have a walk-forward optimization logic that automatically runs when a specific condition triggers. In another words, it's an self-adjusting logic. I can have 2 systems, that has no common categories from the list above.
Now, on multiple occasions I observe that these 2 different systems adjusts closely. Based on the observation, I run a test that triggers the opt. logic for both systems when one of the system triggers. I got a very positive result. So I have a "system of a system" that re-adjusts it's parameters in multiple systems triggered by other systems.
If I was able to extract the specific tendeny of why the 2 systems are connected, I would have done so and categorized it. But I don't have a logical explanation and I don't have a statistical measurement to detect these kinds of "relationships". All I can say is, it adds value to my trading.
So I'll end this post with another cliche...
Everything depends and there's no magic formula.
TEST EVERYTHING.
PS. Wooohooo!!! I didn't end it with a question. Or should I end it? :eek: ooops...
So, let me get this straight... you have key statistics that you use for the sub-categories that ulimately tell you how to allocate your money. And you have categorically different systems that relate to each other under a certain condition, and you're trying to figure out how to measure and/or categorize them. Is this correct? Do you notice this sort of thing with other different systems or just this example?
Why not categorize your functions by an attribute of times
(shocks in security types,market volatility, economic conditions, interest rates,correlation to portfolio, contango/backwardation, yieldcurve,etc.) ? If you're trading off of these signals without an explanation, then you could create variables for the relationships.
Out of curiosity... how do you compare the categories-- what do you do after you have your grand measure?