Curious others thoughts on combining systems and optimization. When developing a system, often the parameters are varied in search of a set yielding results suited to the designers tastes. Best profit, highest MAR, etc etc.
When adding a second, or subsequent system to the portfolio or suite, of systems. Should the multiple, two for example, systems be paired and both traded at their individual 'best' parameter set. Or, would it be preferable to optimize both systems simultaneously, and select the best symbiotic combination?
I suspect that each approach has its benefits and drawbacks. On the plus side, all in optimization would uncover synergies between the two methods and exploit them, creating a new hybrid system of sorts. This benefit could also become a negative, if all we have accomplished is to make a superior fit to the past data at the cost of reliability.
Backtesting alone, or all in (things that make you go, hmm)
Maybe you would be most comfortable if you could be confident and proud of each individual trade, all by itself, separate from all the other trades of all the other systems. Maybe you don't want to hold your nose when you take trades from system S3 because you believe "S3 is a real stinker but it seems to help the suite somehow." Maybe when your suite is in a bad drawdown and you're in a nasty mood, you would be greatly tempted to skip the trades from system S3 because you think it is a stinker. Maybe the thing to do is avoid this psychological drama, by never including stinker systems in your suite.
Maybe one way to avoid stinker systems is to avoid "co-optimization" entirely (as you suggest).
Maybe another way to avoid stinker systems is to permit "co-optimization" but reject solutions that move any one system more than X% away from its standalone optimum. Where X% is chosen by consulting with your own gut. Perhaps one rule you might consider is, "I will accept co-optimized suites as long as the final co-optimized parameter set does not decrease any individual system's RobustSharpe ratio more than 20% below its standalone optimum value."
Or maybe you don't have any psychological hangups about trading stinker systems at all. Maybe you violently disagree with the very idea of "stinker systems"; maybe you think of them as "valuable sources of anti-correlation." In which case, maybe you think co-optimization is the cat's pyjamas.
Maybe one way to avoid stinker systems is to avoid "co-optimization" entirely (as you suggest).
Maybe another way to avoid stinker systems is to permit "co-optimization" but reject solutions that move any one system more than X% away from its standalone optimum. Where X% is chosen by consulting with your own gut. Perhaps one rule you might consider is, "I will accept co-optimized suites as long as the final co-optimized parameter set does not decrease any individual system's RobustSharpe ratio more than 20% below its standalone optimum value."
Or maybe you don't have any psychological hangups about trading stinker systems at all. Maybe you violently disagree with the very idea of "stinker systems"; maybe you think of them as "valuable sources of anti-correlation." In which case, maybe you think co-optimization is the cat's pyjamas.
My personal take on this owes a lot to the work of Ralph Vince, and can be briefly summarized as follows: 1) use simple robust systems that have a positive mathematical expectation and exhibit a low variance between trades then 2) use a money management strategy to leverage the results.
I have posted previously on how I measure goodness of systems.
Sluggo mentions correlation. Personally I am yet to be convinced of the value of correlation for these 2 reasons: a) it is neither consistent or reliable; b) when the poop meets the fan everything becomes correlated.
Now, to your question. I much prefer systems that exhibit goodness over a wide parameter range so that optimization for "best" parms becomes irrelevant. I also prefer money management strategies that work well over a wide range of parameters. IMHO optimization is only useful in development and testing in order to see where the "sides of the hill" are.
I have posted previously on how I measure goodness of systems.
Sluggo mentions correlation. Personally I am yet to be convinced of the value of correlation for these 2 reasons: a) it is neither consistent or reliable; b) when the poop meets the fan everything becomes correlated.
Now, to your question. I much prefer systems that exhibit goodness over a wide parameter range so that optimization for "best" parms becomes irrelevant. I also prefer money management strategies that work well over a wide range of parameters. IMHO optimization is only useful in development and testing in order to see where the "sides of the hill" are.
Good find, that... Ill copy the relevant paragraph here from the post by Forum Management:levijean wrote:viewtopic.php?p=10697&highlight=#10697
"Another, even less intuitive truth (and one the better of my fellow fund managers might wish I/we didn't discuss) is that a portfolio of systems that have been opimized to have the best performance characteristics in isolation as individual systems, most often does not perform as well as a portfolio of systems that have been optimized to maximize their noncorrelation with each other. " f.m.
So the mid 18% grey zone Sluggo mentions comes forward. Optimize, but within reasonable bounds from the singular best set. Or, Is F.M. suggesting to bend the lines to the point where prudence would suggest angels not tread. hmmmmm Perhaps its the cats pyjamas...
I agree with you regarding Vinces' guidelines! I suppose when systems are of a similar nature, that the most robust spot on the curve as individuals, would seem to be the way to go. Maintaining their best spots as similar timeframe and orientation contributors. (as Sluggo maybe mentioned at the outset)Roscoe wrote:My personal take on this owes a lot to the work of Ralph Vince, and can be briefly summarized as follows: 1) use simple robust systems that have a positive mathematical expectation and exhibit a low variance between trades then 2) use a money management strategy to leverage the results.
I have posted previously on how I measure goodness of systems.
Sluggo mentions correlation. Personally I am yet to be convinced of the value of correlation for these 2 reasons: a) it is neither consistent or reliable; b) when the poop meets the fan everything becomes correlated.
Now, to your question. I much prefer systems that exhibit goodness over a wide parameter range so that optimization for "best" parms becomes irrelevant. I also prefer money management strategies that work well over a wide range of parameters. IMHO optimization is only useful in development and testing in order to see where the "sides of the hill" are.
I suspect that when the nature of the systems becomes more dissimilar in time and target that all in optimization comes into stronger play. For example, a shorter term system or counter trend system combined with that long wave riding board. {surfer talk there) In this realm, tuning the pair in effect creates an entirely new hybrid approach. (one which could probably be duplicated in one singular super-system.) The goal still being to find that cozy spot on the hill.
So, less is more until more is less.
