Page 1 of 1

artificial portfolizer

Posted: Thu Jun 08, 2006 11:03 am
by michal.kreslik
Hello,

the goal of building a trading systems/markets portfolio is to decrease the overall drawdown and to smooth out the overall equity curve. Obviously, this can be done only if the portfolio elements are mutually uncorrelated (cor. coeff. = 0) or better, negatively correlated (-1).

But what if a sound trading system itself is versatile enough to provide uncorrelated or even negatively correlated equity curves according to the parameter sets used?

I am going to test how will the same model behave on the same set of markets if the only "portfolizer" :) is a different parameter set.

Is anyone here using this technique so that I may save time reinventing the wheel?

Thanks and have a very nice day!

Posted: Fri Jun 16, 2006 12:07 am
by eudamonia
Michal,

Um, actually negatively correlated systems -1 are actually perfectly related to each other but in opposite directions. Additionally, just because your portfolio elements have a low correlation (i.e. close to 0) does not mean that your portfolio is not suceptible to fat tails.

I have a rather extensive e-book on Portoflio Iceburg risk. Lots of great math examples in here and the problems with using just the normal distribution curve to make your assumptions. Its too big to post on the site, but let me know if you'd like me to e-mail it to you.

Edward

Posted: Fri Jun 16, 2006 12:20 am
by michal.kreslik
Speaking of correlation, we should put things into perspective. Do we talk about the equity curve correlation or the trade generation correlation? Obviously, if the trades were generated in the same market with the -1 correlation and one of the systems was profitable, the other one would be a loser. Thus, we would not be interested in such a system at all.

Posted: Fri Jun 16, 2006 12:50 am
by eudamonia
Good point. Its important to know which set of correlations we're analyzing. Personally, I believe trade generation is the more appropriate correlation to analyze in this instance. Otherwise a loosely correlated equity curve correlation between two different systems might or might not be well correlated. For example if you compared two trading systems and each were identical but one only traded half of the time, it is quite possible that you would not get a highly correlated equity curve. However, your actual correlation of trades would be quite high.

Posted: Fri Jun 16, 2006 2:58 pm
by jhtumblin
Well I have actually been interested in this idea for some time and have applied it to many of my biggest "losing" systems. Note I am talking about trade correlation here and not the equity curve.

It is my opinion that if I create a strategy that produces a profit factor of 0, it is just as good as a strategy that produces a profit factor of 100, for all I need to do is reverse the trades or the formula that generates the trade in order to move into the profit column. I wouldn't be surprised if many people overlooked this fact, created what could have easily been a winning strategy, and then discarded it due to it's terrible (numbers) in backtesting.

Posted: Fri Jun 16, 2006 11:10 pm
by eudamonia
Not necessarily. Execution costs are still king. Often the reason that a system loses isn't that it doesn't have any validity whatsoever, but rather because execution costs and too many trades make the system untenable.

Posted: Mon Jun 19, 2006 2:43 pm
by michal.kreslik
As a reply to your suggestions, I've started a new topic on the theme of assessing the difference between the healthy system and a tradable one:

fighting the randomness: healthy system vs. tradable system