Enabling / disable of systems based on performance?

[Oppz]

One thing that I have started to think about is if there would be any advantages to having additional filters for trading a system of systems.

Let’s say that I have 20 systems / EA (different currencies, timeframes, parameters etc. ), but in total 20 systems that are reasonably uncorrelated. I guess the normal way most trade this is to have all 20 enabled, but what if there were additional filters on this, like that only those systems that are above a MA (or any other suitable parameter) are traded? The tracking of the performance must of course be made offline (backtesting, paper trading or any other method), so trades are tracked even if the actual system does not place any orders . The result may be that only 15 are traded at one point, 20 at another, and 10 at a third time.

The major benefit would also be that you safeguard the system you have if it no longer works => no trading.

But I can’t find anything about this… You are basically trading a system of systems, activating different systems based on their performance / the equity line but on a system level. I would sure be a question more people then myself would have askt themself. But I cant find anything about this… but I may search in the wrong places (or wrong terminology).

Any thoughts about this?

 

[globalarbtrader]

This only works if system performance is positively autocorrelated, and the switching on and off can overcome the cost penalty you will be be paying.

More here

GAT

 

[Oppz]

This only works if system performance is positively autocorrelated, and the switching on and off can overcome the cost penalty you will be be paying.

More here

GAT

Thanks for the link, exactly what I was looking for. Interesting to read the article and other articles he links to. It seems to be an area with not so much analysis/articles out there (as the author also writes).

I did only read it briefly for now, but there were some areas I didn't really follow him (or directly agreed...).
1. He seems to apply this to only 1 EQ/System. The point in my view would be to apply this to a system of systems.
2. Random data / EQ. I understand why he uses that, but not sure of the benefit of this, especially with 1).
3. Seems to be using way too short periods, realize that frequent on / off will create problems. My initial idea was to use something like 6months rate of change (or similar).

But he sure knows statistics way better than I do, so I will read it again and all links. But highly appreciated, and all links about this are welcomed. I have read Rob Carvers books, and he is fore sure knows what he talks about.

 

[globalarbtrader]

Thanks for the link, exactly what I was looking for. Interesting to read the article and other articles he links to. It seems to be an area with not so much analysis/articles out there (as the author also writes).

I did only read it briefly for now, but there were some areas I didn't really follow him (or directly agreed...).
1. He seems to apply this to only 1 EQ/System. The point in my view would be to apply this to a system of systems.
2. Random data / EQ. I understand why he uses that, but not sure of the benefit of this, especially with 1).
3. Seems to be using way too short periods, realize that frequent on / off will create problems. My initial idea was to use something like 6months rate of change (or similar).

But he sure knows statistics way better than I do, so I will read it again and all links. But highly appreciated, and all links about this are welcomed.

By the way, I am the author of that blog post

GAT

 

[Oppz]

By the way, I am the author of that blog post

GAT

Well... Nice to meet you and thanks for all your effort and the books you have made, really good trading books.

So what's your thought about my "concerns" for 1 & 3?

I realize that a system-of-system approach never will be better than a bunch of really nice systems in terms of profit, but one other huge benefit is to have a fail-safe if one system stops working due to whatever reason. That's why my initial idea was to have such a long (slow) filter, to let the everyday noise play out, and just safeguard the major drawdowns (or stop to work). And the more systems I use, the bigger of course the risk that this will happen.

I may have completely misunderstood your blog post (and apology if so), but isn't the result basically the same as if the EQ line would have corresponded to a Buy&Hold strategy of a random instrument? But then again, it was 20yrs since I did any university statistics...

 

[globalarbtrader]

Well... Nice to meet you and thanks for all your effort and the books you have made, really good trading books..

Thank you.

So what's your thought about my "concerns" for 1 & 3?

1. Results would be the same for a set of systems - unless the returns for any given system are positively autocorrelated, then turning some of them off will not be a good idea.
2. I use random data to avoid the chance that the results I find are just pure fluke and to understand the properties of a trading strategy that are important
3. " Seems to be using way too short periods" I use everything from two weeks up to two years.

I realize that a system-of-system approach never will be better than a bunch of really nice systems in terms of profit, .

NO. A system of systems is much better than any individual system, because of the likely diversification benefits (I have ~50 systems running on ~150 futures markets). Don't let my dislike for switching systems on and off make you think less of this idea. What I am saying is that it's really hard to systematically turn systems on and off in such a way that will make you more money than having a system of systems.

I may have completely misunderstood your blog post (and apology if so), but isn't the result basically the same as if the EQ line would have corresponded to a Buy&Hold strategy of a random instrument? But then again, it was 20yrs since I did any university statistics...

No

GAT

 

[Oppz]

I did additional testing, and the shocking conclusion is that it is "better" not to use this as a filter, as both the profit and smoothness of the equity curve in my backtested simulations is higher / smoother without this filter.

As how.... I will still use a simple 6-month filter for the system, not to get a higher return or smother EQ, but simply as a safeguard if a system starts to fail. Maybe that's the major take from this.

 

[globalarbtrader]

I did additional testing, and the shocking conclusion is that it is "better" not to use this as a filter, as both the profit and smoothness of the equity curve in my backtested simulations is higher / smoother without this filter.

As how.... I will still use a simple 6-month filter for the system, not to get a higher return or smother EQ, but simply as a safeguard if a system starts to fail. Maybe that's the major take from this.

How over just 6 months can you distinguish between 'failure' and just random noise.

 

[Oppz]

Good question, It's based on old-fashioned backtesting. A 6-month filter will, based on 20yrs of data, only have turned the systems "off" a few times. There is of course always a risk, but based on history I think I can live with it.