Monte Carlo

Hello,

I was wondering if someone could kindly share their approach to using Monte Carlo within SQ. I currently have 10,000 EURUSD strategies and I want to decrease this amount down to the top 100 and create some portfolios from there.

I was thinking to repeat-apply Monte Carlo until all 9,900 strategies have been eliminated, what do you think?

Is there a standard number of iterations that people are using – I read that 10,000 iterations gives something like a 95% confidence level. Of course, we cannot have a 100% confidence level, but I am just looking to get a close as the program will allow, while still leaving me some breathing room and not eliminating everything.

My start point would be for 10,000 strategies – use 100 iterations to reduce the number; gradually increase the iterations (to save time, decrease the sample quicker and ensure I don’t overshoot and eliminate everything)

Thank you,

Matthew

I use most times 200-500 iterations for the montecarlo

did you already made some not so time consuming tests – like higher/lower TF, another market, etc.

because running MC tests on 10000 strs is not reasonable

how did you get those strategies, did you try another settings, or this is only one setting at all?

did you have only test on all EURUSD data?

Hi Matthew,

I agree with Hankeys. Once you have loads of strategies created in the builder mode, it is a good idea to start with “fast” robustness tests in the retester and work from top to bottom. It is best to only work with the slower tests once you already removed the majority of your strategies with the faster tests. See also the tutorials of SQ.

Hi,

Thanks, I am not going to test my currency pair just on Monte Carlo – I also use

1. Retest with different spread / slippage

2. Retest with Higher Precision

3. Retest with different timeframes and currencies

Since those 3 tests are not random, I thought Monte Carlo should come at the end of the testing cycle. Since Monte Carlo is random, it can give different results each time you run it, so I think its useful to reduce a number of tests which also pass the above.

Do you agree about the order? Maybe there is no “right answer” but Im thinking there should be a uniform approach.

In answer to your question, these are randomly generated with some loose ranking criteria (NP; RDD; Stability; Stagnation, #trades) etc. I tend not to change the settings (you mean spread / % traded etc?) during the generation process since I found that can adversely affect the retesting since some strats are more outside the range than others.

Thank you, yes – I will look to work from top to bottom – the Monte Carlo tests take a long time, so its better to do those on lesser number of strategies – that was my suspicion and you have confirmed it, thanks 🙂

Do we agree with TNickel that 200-500 iterations are good, or should I keep going until and hitting “RETEST” until my databank has 100 strats in there?

I was reading around the statistical approach and the more iterations there are, which pass, the more robust the test – so I was wondering is there is an optimal number to aim for, in a general sense.

my approach is different – i dont like BIG basic packs of strategies generated with the same settings…

i am using for now custom projects where i run all tasks together and work only with a “final” strategies

for MC tests i am using 200 iterations

I don’t know whether it is optimal or not but I prefer 1000 iterations. I’ve reached this conclusion by exploring with 100 iterations. If 100 iterations are employed, the remaining strategies tend to change for each MC run. When 1000 iterations are employed same strategies started to pass MC test regardless how many times the test conducted.

Interesting, thank you !

it could depend on what MC test you are doing – for example MC RANDOMIZE TRADES EXACT, there is endless combinations, and you can make this test repeatedly and you will come to a point, that no strategy will left you

its only probabilistic methods with some value of confidence

on the other side – MC test change parameters, where you set the probability  and max change – so you are limiting the range where the MC tests will be done

making 1000 iterations on MC test change params is waste of time by my opinion

also MC test spread and slippage, there you could very well live with 20 iterations only, because the range of values is not huge…its nonsense to make 200 iterationg for slippage 1-5 for example

Thanks everyone. I am not sure yet how to use the workflow tools but will get around to that when I have a full license.