Forum Replies Created
-
Posts
-
Hi Steve,
First point is that you can never be 100% sure that a generated strategy hasn’t been over optimized or will perform well in the future really. Robustness testing will certainly reduce this likelihood and I wouldn’t be trading a strategy that didn’t pass this filter.
It sounds like you are generating a collection using the reactor and then adding them all to the portfolio and then checking OOS and not getting a profitable single asset portfolio? This isn’t too surprising really
I would be putting together a multi asset portfolio rather than a single asset portfolio.
Hey Alan that is all sounding good and yes now the next step is to see how many reactors is optimal for each machine. I am guessing 3-4.
The number generated is the most important bit here, I’m sure if you left the above test running longer then difference in the number collected gap would widen but then also as you get more strategies the likelihood of correlated strategies not making it through increases. Main point is all is looking well now and you are fine tuning and starting to get to work on your personal pool of robots :-)
Yep Alan, that’s the way to use the validator to organize your collections and yes also combining in this way will be the best point to sort and filter the final 10.
For general robustness the Monte Carlo is enough.
Unless you are trying to target something specific with the Multimarket or Walk forward validations but otherwise Id just stick to Monte Carlo. That’s what I do.
It sounds like you are building up your pool of strategies nicely now right?
Yes combining them by running through validator is a good idea. Mainly because it will remove correlations and duplicate strategies and also it’s good to run the Monte Carlo a second time. But even if not running a second Monte Carlo and just running through the acceptance criteria the validator will trim out the correlated strategies that wouldn’t have been collected had they all been generated in a single reactor.
Hey Alan,
If you are using the Monte Carlo on the validator then that is what is going on. There is randomness in the Monte Carlo sometimes s strategy may just pass the 80% threshold and sometimes it may just fail it for example. But that’s okay the 34 going through are only the ones that have passed the Monte Carlo twice :-)
That all sounds great Alan, you are well on your way now :-).
Glad we got it resolved eventually. Was a puzzler for a while there!
In my experience each extra reactor will slow down things a little bit but there is a sweet spot, for example adding a second might mean that each reactor generates 1000 a minute which means combined you are getting 2000 a minute, a third might reduce this the each getting 800 a minute which is still 2400 combined so an over all improvement but then suddenly a 4th might slow it down more considerably to only getting 500 a minute and then thats 2000 combined and so the sweet spot is 3 reactors.
Hey Alan, that’s definitely sounding more reasonable and giving you something to work with.
Make sure you save all your work and build up a nice pool of robots so that you can save yourself a lot of time in the future by reusing your results with the validator :-)
Also am I right in thinking that you have already purchased petkos Ea package and get robots every month?
You run the reactor to generate from scratch but you can also load in as many existing EAs as you can to the validator and run with the same settings, that way you will get lots of EAs into your collection MUCH faster.
Also make sure you check how running a few reactors at same time on each laptop goes as I’m pretty sure that would be the optimal amount for your hardware (6 reactors in total is absolutely plenty), and always keep an eye on those generations rather than only what’s collected and make sure that after 10 hrs you have equivalent of roughly 1000 a minute for 10 hrs straight on each reactor
Sounds like all is well now though and you just need to get to work in building up a nice pool of robots for yourself now right? ;-)
Yes Alan, that is exactly what will make the difference! If you put another window in the front the browser will become inactive. This is the cause, we got there eventually :-)
You need the screen on and can have several browsers running in the foreground as long as you position them to all remain active.
There are some cost effective vps options but if it doesn’t feel like an preferred option then totally not required, it’s just good to have options :-)
Yeah I agree Petko, EA studio code is incredibly fast and doesn’t require much hardware but current issues could make it a good option. If prevent screen turning off then there shouldn’t be an issue with reactor going inactive though!
The main issue here seems to be somehow the reactor going inactive, a vps can be an option worth pursuing for some. I’d probably do it if I didn’t have my desktops just to allow me not to have a screen on all night.
Kind of a preference thing really though.
Hey Matthew,
You are definitely working things out and seem well on your way to that ftmo prize :-).
Its great to the the good results and I definitely agree. If it ain’t broke don’t fix it, sticking with a system is a major key to success. I got lost a while back jumping around and trying too many things too often.
keep it simple and if it ain’t broke don’t fix it are two excellent rules to live by in trading.
will be interesting to see your results when you get more pairs involved for sure!
Hey Alan,
Wel that aligns with me getting about 40 in 10 hrs which is good. Remember and keep an eye on number of generations though. Also definitely try and get a few running at the same time on each computer and check that it doesn’t slow the reactors too much.
re vps well that depends on how many reactors you want to run at the same time. I would think that with your two laptops you should be able to run 3-4 reactors simultaneously one each so that may be enough for you (it’s plenty really).
I wouldn’t get anything fancy.
You probably have enough power without but if you did want to go that route then being able to run 4-8 reactors 24/7 is plenty in my opinion so a 4-6 core would be a good choice.
Determine how you want to use EA studio etc first before deciding what you need.
You do probably have enough computing power already if you are happy to leave them on 24/7 but I guess I am just thinking screens draw a lot of unnecessary power in this case so possibly not a very elegant long term solution.
but you could maybe just opt for one and try it for a month and see how it suits you, can cancel at any point?
I think we might be close to solving it but yes let’s not count our chickens yet. However what I am seeing is that your 10hr reactor generated just 20% more than what you have just shown was done in 45 minutes. So what is happening is your reactor is going inactive after about an hour and therefore your 10hr run is really just results for an hours run.
So if you were getting typically 3 you could expect about 30 if it had actually run all night.
So before we worry too much about the number getting through the acceptance criteria and the Monte Carlo we want to see that 600,000+ generations after an overnight run.
Does your screen stay on all night? What would be a good option is to get a vps for running EA studio on. I don’t get these issue because I run on desktop.
Also getting half what I get on my desktop is probably fair and should still allow you to get plenty of strategies. How many cores does your processor have?
Also 10% passing Monte Carlo isn’t necessarily a problem, we use it to leave only the more robust strategies and sometimes that is just about 10%. Usually I like to see around 10-25% getting through the Monte carlo
You can run Ea studio on as many computers as you want Alan. I have two mid range desktops running about 6-8 reactors on each computer and each reactor does about 2000 generations a minute which would be about 1.2 million over 10hrs so yes something is definitely up with your computer speed as it is running 22x slower than mine?
For reference my processors are just ryzen 4700g on one machine and intel i5-10400 on the other so really nothing fancy at all both machines cost me under 1000 AUD about 18 months ago.
Let me know how you get on but looks like we have found the issue at last ;-)
Hey Steve,
I’ll take a second look at your post tomorrow as there is a lot to digest in what you are trying to do.
A couple of initial points though.
One is that when Petko suggests checking with 200k bars if have generated using 100k bars this allows a check of OOS BEFORE the insample data as opposed to after the OOS which is done by setting OOS % in generator.
I agree also to stick with Monte Carlo only for robustness in the Reactor. You could increase it’s effectiveness further by increasing the number of tests, or afterwards running it through the validator a second time perhaps but for me it’s the best robustness tool.
I also never really using the optimiser but in this case I can see how it probably is useful for what you are doing. As you have so few indicators and no exit indicator along with not allowing optimization of stop loss and take profit I think it’s okay and has a much lower risk of over optimization than normal but generally I would avoid.
In this specific case I would personally either use only the optimiser in the reactor as you have the setting OR use the normalizer one the reactor as Petko suggests and then for each strategy optimiser only the time entry indicator with reasonable size steps (The more time consuming option but the better of the two in my opinion).
So ideally in a nutshell: Reactor with normalizer and Monte Carlo and then optimize only the time setting on an individual basis with hourly or two hourly steps for example.
Also check out the time setting in tools for either control of when your strategy trades.
Hope that helps
Hi Alan,
Yes that is what Petko is saying and it is definitely the case.
Make sure that you close other unneeded applications and browsers when running the reactor. Also take note of the number of strategies generated and not just the number making it into the collection if you’s want to gauge how optimally EA studio is working.
My rough rule of thumb is to run a separate window running a single instance of EA studio per processor core.
Also make sure that your browsers are active, you cannot minimize them. This is the main reason I want to see a screenshot actually, I want to see how many strategies have been generated rather than added to the collection. I suspect the window may be going inactive.
After an overnight run the number of generated strategies should be many hundreds of thousands and maybe even over a million depending on computing power.
Hopefully that’s the issue as I really don’t think it’s to do with the account settings like I originally suspected.
Oh and always use chrome of course :-)