Data Window size
14 replies
RNG
4 years ago #241802
Hi guys, i’m here to open a productive topic.
I’m wondering why Mark and the developer team haven’t put a main feature yet.
I’m talking about the automatic data windows size.
Like you know you can’t choice a random windows size, or at least yes you can but this can be unproductive. I’m explain better myself, when you start a process of building, we are searching for productive/robust strategies in the nearest future, and of course for example, if you are searching H1 strategies that are working in the next month, is totally useless use 12 years data old, don’t you think? Perhaps maybe not everyone know, there is a scientific way to spot the right windows size based on your purpose.
So here it is, before to write it down, i wanna hear some community thought about it, how do you choose the windows size? Do you know the scientific method?
This is for you Mark, do you know it? Do you and your team know how to calculate the right windows size of the data?
After a little bit of comments i’ll explain how to do it.
4 years ago #241805
I wanted to know more about it.
mabi
4 years ago #241820
Seems to me that the more data a strategy worked on in the passed the more likely it is to work on data going forward. Since you have the possebility to make plenty such strategies it would be better to just retest them on recent data to see which ones is working now and use those. I think you will find that the % that works will then be much higher then just strategies made on a recent window.
RNG
4 years ago #241826
Not totally correct, yes of course, more a strategy is working during the History and more is reliable through different kind of markets, but we aren’t searching for one “holy grail” strategy working well always and working forever.
A strategy created on 12 years or more, will be a very slow strategy, and will miss a lot of gain opportunities. And consequence of the big data window size is an overfitting of the builded strategies, to aim a good return/DD during all this time SQ will create a very complex rule/rules of entry exit, managing etc.
IMHO I prefer create strategies more frequently, and optimising/deleting and swapping it when they lose the reliability by monitoring all the stats.
It’s weird that this topic have so low success, i was thinking a more light on by the community…
mabi
4 years ago #241830
This topic is difficult. For example make strategies on last years data but then exclude this recent may. You find that 95 % is looser in may well depending on how you rank them . I did another test as well yesterday, Tested 3800 EU that was really good during 2018 rdd >5 ( they were from a batch of 25000). 700 of these 3800 worked great month Jan-march the rest was losers . 525 of these worked great in April 2019 80%. But only 3 of these 525 was profitable in may the rest really big losers. So then I tested all 25000 on may 2019 they are made and tested on data from 1986-2018-09. Now I had about 2700 that was profitable in may 2019 . Now I went further and test them month by month for no loosing months ( but flat is okey) going back 12 months. I have 37 left they do not have an RDD above 5 on 2018 so they cant be from my original test of the 3800.
So maybe that a good workflow. I know have 37 strategies that worked great the last 12 months but also ok for the last 30 + years.
Attached is how they look on Duka.
clonex / Ivan Hudec
4 years ago #241832
“Do you know the scientific method?” Tell us
RNG
4 years ago #241835
Clonex i’ll tell you, but before i wanna listen some different point of view about it.
RNG
4 years ago #241878
Dead end for this topic? For example on EURUSD H1 in this moment there is a valid window
From: “10 October 2018”
To: “26 April 2019”
Tradable in two next windows:
Window 1: From: 27 April 2019 to 21 May 2019 with accuracy of 90%
Window 2: From: 22 May 2019 to 27 June 2019 with accuracy of 80%
No one guess how i know this?
C’mon guys be active!
mabi
4 years ago #241887
I have no clue. Only thing i know is that You are referring to History. I suppose it is easy to prove later by testing performance on a couple of hundred live strategies on Your suggested history data
mabi
4 years ago #241890
Hopefully You are right thought 🙂
Gianfranco
4 years ago #241898
hi …Gianfranco hurst exponent….for scan symbols and data Windows….up 0.5 ?
is my little idea….automatic scan with min and max days…like WFM
thanks
Gianfranco
RNG
4 years ago #241903
Someone say it, finally!!!
Yes is the Hurst Exponent, we have a lot of statistic tools to be sure of working properly.
Now i explain my problem, i have a lot of python scripts: to clear the data, find the data window, correlation between data, manage the building, find the uncorrelated strategies, monitoring the equity and chose for reoptimization or deleting.
There are a lot of trick to speed up the entire process, for example, correlation of strategies with a very big pool, 10k strategies, instead of testing entirely with all the data window, i’m shrinking the data window to 1/3 (IS-OOS) of the original, and using those equities, less computation power, split the pool in two parts (good correlations between -0.5 and 0.5, and bad correlations between -1 and -0.51, 0.51 to 1 retesting the correlations of goods with the original data size to be sure of the partial results, and that’s it, a lot of energy and time saved.
I’m asking myself why they didn’t include yet all this tools inside the suite?
However if someone need those scripts i’ll share it, but i prefer share it with the development team to join it inside the software.
hankeys
4 years ago #241905
yes, if you are or have a good programmer many things could be much more easier….we have for example python script for old SQ to control MC tests, etc.
You want to be a profitable algotrader? We started using StrateQuant software in early 2014. For now we have a very big knowhow for building EAs for every possible types of markets. We share this knowhow, apps, tools and also all final strategies with real traders. If you want to join us, fill in the FORM.
hankeys
4 years ago #241906
i have read some article for these hypotesis – https://www.mql5.com/en/articles/2930
but i dont know how to use them in our “bussiness”
You want to be a profitable algotrader? We started using StrateQuant software in early 2014. For now we have a very big knowhow for building EAs for every possible types of markets. We share this knowhow, apps, tools and also all final strategies with real traders. If you want to join us, fill in the FORM.
mabi
4 years ago #241919
I do not belive this would work using timeseries trying to predict short future periods in FX which is totally randome. If it works one period it was only luck. I rather Optimize an entry over 100 years of data to get an good average performance and then by walfkforward try to adapt the exits instead to current market. This way you can also get strategies with >90% winning 6 months periods looking back 30 years and automate the workflow with current tools.
Viewing 14 replies - 1 through 14 (of 14 total)