|
|
Owner/Admin
Posts: 3922
   
Location:
USA: GA, Lawrenceville | Mark and Steve have done a lot of ECA test runs, which start by using a pre-filtered limited-number set of good strats to produce portfolios with 7-14 strats active at a time (or thereabouts, as I understand it).
Keith asked (click here) whether it might be better to simply throw ALL the possible strats (note this is a theoretically infinite number but let's just think of it as a whole buncha lotta strats) into ECA at once, with lots of eval rules, then push the button and, after a long wait, out pops the ideal solution.
Aside from the fact that it would simply be "too easy, Drill Sargent" - and would take a truly long time to recalc, apparently M+S's testing (which Ed presented in a recent webinar) shows that the results of doing that kind of thing just don't work out to be better - versus starting with a limited slew tied well-thought out set of strats.
Here's a "theory" that might have some relevance, and help explain this.
If you start with a huge number of strats, amongst which many have only minor variations from one another, and let ECA do it's thing, then theres a strong likelihood that "curve fitting" might result *in the sense that* for a given monthly Walkforward step, ECA would "discover" that on particular Strat (plus all it's close cousins) really shines. So, for that WF step, ECA would "stack the deck" for the following month with that collection of closely related strats.
Then, presume that the next month's characteristics differ enough from the prior one that the behaviour of the "stacked deck" of strategy-cousins all simultaneously get tripped up. In that case, even though we are using multiple strats, we'd see a magnified drawdown.
The "close-cousin" relationship might occur for several reasons - it could be several strats built on the same System with differing lists or conditions - or could be somewhat similar systems (ie all "short term trending") built on the same fairly short filtered list - or could be even related to some kind of overtuned condition that forces the similarity. Also, short symbol lists increase chances that larger positions are taken on one symbol (if acct settings permit).
| |
| | |
Owner/Admin
Posts: 3922
Location:
USA: GA, Lawrenceville | How to remedy this (admittedly theoretical) situation?
"Intelligent Design".
That is, deliberately limit the collection of strats that are considered for the ECA hopper to ones that are diverse in their performance characteristics.
Or, focus the strats in a portfolio deliberately to work with some specific automated-definition MarketState, which might include Strat conditions &/or tailored dynamic lists for the MarketState. Then, use PortfolioBalancer to auto-switch between the MarketState specific port's.
Alternatively, you could deliberately combine Strats that are individually tailored (via conditions &/or MarketState-specific dynamic lists) into an "eclectic" portfolio which in theory would have enough "goodies" in it to work well with many different MarketStates.
In the TradeTight forum, our focus will be primarily on the creative development of Conditions and Dynamic Lists for this kind of use, as well as discussions about *why* admixtures seem to work better than others, so that users can use their time wisely to develop and tailor their own OVest accounts. | |
| | |
Veteran
Posts: 212
Location:
USA: TX, Mansfield | I have been dipping my toe in the water lately but haven't come close to the results I see being posted. What I don't see posted are the strats being used (just might not be looking in the right places).
It would be helpful if folks would share the specifics of how they build their runs for those of us who are lagging behind and need a little hand holding.
Don | |
| |
|
How Many Strats s/b input to ECA?