Quote: Originally posted by RandyT  |
@getty, a few comments to your questions.
1. I would be very surprised if you could find 12 uncorrelated systems to trade on the same market. I would love to see you prove me wrong (and share
how you achieve that) but I think this is a difficult task. GSB tends to find breakout systems in my experience. Mean reversion for example is not
something that can be produced in GSB currently.
2. There is no way to do this currently in GSB. Also, in my experience, it is difficult to find systems over last 10 years that are not flat some part
of it. It is all shades of gray as to how much better some systems do over those market regime changes than others.
3. A GSB run has no way of knowing what the performance is of another system you have in your portfolio. Interesting idea, but seems a challenging
task. I think this will need to be part of your workflow with PA assessing your results after runs.
FWIW
Edit: Just to add some context to perhaps help with expectations. As an example, the work that Peter has been sharing on CL system development for
example, has been months of work. It takes a ton of system resources and lots of time to try the huge number of different possible configurations to
reach best system performance. Not meant to be discouraging, but rather to help prepare for a lot of work ahead. Would highly recommend absorbing
Peter's latest work on CL as that is the current "state-of-the-art" on system development. |
Thanks very much for the reply Randy. It's great to share information on such things as my experience has been that many strategy developers don't
want to discuss *anything* for fear of losing alpha. I much prefer this GSB community approach of openness to collectively conquer these really
challenging problems as a group.
I realized after posting, just how unlikely it will be to find 12 decorrelated strategies on a single symbol. But having said this, finding perhaps
up to 6 strategies on a symbol isn't unreasonable by simply exploiting different market regimes - trending, bear market, momentum, volatile, etc. By
employing a regime approach I could expect to have very little correlation between strategies. This would also address my question 2, by identifying
strategies that work during low volatility. I've already requested with Peter the ability to perform regime switching.
In a sense, GSB is focused on the volatility regime which is why long periods of the ES system, for example, are flat. The search capabilities of GSB
are very powerful. if I were to speculate, part of the reason that low volatility challenges GSB is the frequent lack of use of daily data where
signal to noise is high compared to intraday. Yes, volatility creates high SNR which generates us better alpha. But in lower volatility (and lower
SNR), longer time-frames are generally needed to become successful (hence the request to include daily data).
I've been watching Peter's last 2 videos on continuous loop and reading the community comments in the private area. Perhaps the easiest approach for
my future GSB portfolio to achieve the decorrelated returns I'm looking for is do what I believe the rest of you are doing - invest months searching
the equity and energy markets for my 6 symbols and find my 12 decorrelated strategies that way, rather than regimes.
Regarding 3, one approach I've used before for a correlation metric is to average all values by row in the correlation matrix. The lower value, the
lower the average correlation to all other strategies. It would be possible to put the average correlation into one of the GSB columns, then I could
sort all strategies from least to most correlated. It's possible this may be best done family-to-family. I would see this as an advantage of
performing inside GSB rather than PA Pro because I could sort and remove strategies using a macro (or manually) which are too correlated with each
other.
|
.. then i just wait until Peter has finished the new feature fix for this... Regards