Genetic evolution: Should you recycle strategies as an input population?
7 replies
gottogethelp
6 years ago #257423
Hi all. Once you’ve run a workflow end to end then you’ve built up a set of failed strategies at each stage of the workflow (and hopefully some successful ones at the very end!). Would it be a good idea to collect up a decent number of the later workflow strategies and use them as a genetic evolution input population for the next workflow? Would this perhaps result in too many strategies that are correlated with one another?
0
Oliver
6 years ago #257431
Ive had some success where ive generated some strategies which have not necessary passed all robustness tests. I have then gone back to run the initial OOS test. Those that pass, i then loaded into the initial population tab and found strategies which then go on to pass all robustness tests.
1
gottogethelp
6 years ago #257440
Thanks Oliver. The next stage after Builder in my workflow is to test with additional precision. I found that loading up 400 of the strategies that make it through this first step into the Initial Population increased the Builder speed from 85 to over 600 strategies per hour (and that’s even with an additional stability measure added!). There seems to be about the same level of attrition as I put them through the robustness checks. Haven’t built up enough of a sample to say whether they are highly correlated…
0
Oliver
6 years ago #257455
Yes it does produce a lot of correlated strategies but i find that if you batch process them you can find some really good ones which which pass all the robustness tests and even work better on other markets.
0
gottogethelp
6 years ago #257459
Ok cool. Thanks!
0
kasinath
6 years ago #260522
I have then gone back to run the initial OOS test.
Hi,
I’m new to strategy quant and also looking at how best to recycle strategies.
What do you mean by this statement above? Do you mean that you take the failed strategies and then train them on the OOS data (ie: you use the OOS as IS)?
0
Oliver
6 years ago #260555
so i build using most signals, indicators and price entries using random generation. i then back test those on earlier years. if they meet certain metrics i then keep them and use in the recycling process. the random build process takes a while. i do this because want un corellated strategies. this is the best way i have found to do it without spending too much time
0
Insanity82007
6 years ago #260607
To be honest, you’re better off generating new strategies for 2 reasons:
1. More variety in your portfolios.
2. Less overhead. It’s a headache managing large volumes of pre-generated strategies, especially when SQ can generate new strats so fast. I scrapped 200k strats a while ago for this reason and never looked back
0
Viewing 7 replies - 1 through 7 (of 7 total)