Recommended Initial Data period for Building
14 replies
Andre Alexa
3 years ago #268108
Hello, I am new here. I am building strategies for EURUSD on H1 using default Breakout config.
The data I use for the building is 2011 – 2018 (Eight years) and I set Data range parts with 10 OOS (the last option). I on purpose leave 2019-202 as out of sample for the final OOS test.
I manage to generate promising strategies using the robustness tests as instructed in the training videos.
All final strategies are great when I backtest using long term data such as 2003-Today (17 years) or 2011 – Today (10 years).
However if I run the strategies on 2018-2020 (now) meaning the last 3 years I can see the strategies are struggling. The equity curves are bad. With that fact, seems these strategies are failing to make stable profit for the past 3 years market data.
It seems cuvefitting is still happening even after the strategies have been tested rigorously.
Am I already in the right track? Or am I choosing the data period wrong in the building process?
Should I focus SQX to train on the most recent years (for example 2015-2020)?
Thank you,
Andre
hankeys
3 years ago #268142
truth is, that EURUSD sucks for last years in general
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.
WJPII
3 years ago #268164
Like to chime in here because I also need help/advice on time periods.
I have watched all the videos and read all the manuals for SQX and I am still not able to figure out how to get a somewhat robust EA after months of working on SQX. I would like to know what is the best method for getting a good 10 year back test from an EA. If I start the builder with a 10 year, yes, I will get many EA’s that look good, but will not trade very often. But if I start with a 1 year with same parameters, I can get multiple good EA’s with much more trades, but then it is horrible in 10 years even after optimization. Any advice as to what the best method is for time period of starting builds to moving to optimzation and walk forward tests? Thanks.
hankeys
3 years ago #268166
i am afraid that no one could tell you something like “the best” option – many users many needs and noone knows what strategies are you looking for
i am using always different approaches – sometime i dont use OOS and generating on for example 15 years, sometimes i generate on 3 years, etc. etc.
there are many variables and you need to know what strats you want to find – what SL and TP values, what building blocks, what types (stop/limit/market) – do you want to close on friday or not? what MM will you want to use?
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.
WJPII
3 years ago #268167
Thank you for always helping with advice to my questions. Appreciate your time. I will be more specific.
I have been very specific as to what I am looking for in an EA and have not been successful in finding a way to make it in SQX for over 3 months now. There are EA’s I have purchased in MQL5 that trade the way I like, but I cannot get SQX to get the same type out. Basically, an EA that has the potential to trade daily with stop/limit, has a TP no more than 30 (target) with a minimum TP of 5, BE at 1, SL no more than 80, but minimum of 10. There are a variety of indicator combo’s and signals for the logic on opening and closing. I am successful on getting EA’s in builder on short time periods (3-12 months) and can optimize those until I am blue in the face to get wonderful EA’s. But, then when trying to get a retest and optimization of 10 years, EA sucks. If I start at 10 year in builder then optimization on same indicator/signals, there is hardly no trades and this makes it an unsellable EA because no one wants to have an EA that trades 1 or 2 times a month. In terms of MM, I start with $200 account balance at .01 fixed lots in builder (fast build), then will move to % of balance in retest and optimize.
I think there is also a bug on the trading options portion of SQX on 10+ year optimization with the above parameters set. I will get floaters for 300 days even though I have 1 BE, 5 TP min, 10 SL min, 80 max SL, trade in certain daily times and I still get 200 day floaters. The only way to stop this is to close trades on friday’s and this should not be the case. But that is for a separate thread when I have screenshots.
So, just wondering if people are building on shorter time periods, then scaling up by testing, optimizing, retesting, repeat over and over again by extending the time period each time OR starting with a 10-15 year building. The later is not creating strategies that will trade at least weekly and then obviously seeing these massive floaters.
Thanks again.
Waid
3 years ago #268178
The most popular EAs are set by low TP with high SL in 1 min TF, and most of them are martingale.
Since such settings are more attractive for a retail traders, like what they will do in manual trading.
So to build so, first minimize the TF as low as possible like at least 15min for daily trading.
Also such ‘low TP, high SL’ settings needs high win rates to maintain positive expectation value, so it is a must to filter out low win rates when building (e.g. win rate > 60%).
Low TF means expensive computing power. I am curious about has anyone here ever build a day trading strategy like this?
I think it is not going to work that we only build a strategy base on very thin time period, since you will need to test it on several years.
It is an important topic for AI when deciding training and testing set ratio. It will always going to fail if training/testing is very low, especially you only have less than 100 data set in training set (in our case, IS set).
But in another way that we don’t have a such powerful PC to build 5 min TF based on at least 6-8 years period.
So here is my proposal (although I never try it b4), I call it Jacknife method:
0. build a pipeline by the following:
1. build a 5min strategy based on 6 months or more (less then 2 years)
2. test on an arbitrary year (will small filters like ret/DD > 0.3, # trades > 25…)
3. test on another arbitrary year
4. moving on testing years and years until you have test 10 years
5. the strategies that can pass the above steps is consider as alternate strategies, which pass the whole build of 10 years
then you can move on test on MC or WMF tests and other OOS.
One might argue that such test is equivalent to direct test 10 years in one time.
No, totally different, if you are smart enough, you know why.
Here is what SQ doing:
SQ will test a strategy from 1st to 10th year thoroughly and calculate whether it will pass the filter that you set or not. In this way it is tooooo time consuming for lower TF.
If one strategy cannot pass the 2nd year, then why do you continue? Just throw it away, move on the next year.
So the Jacknife method will build a strategy from cutting the entire ten years into slices, a very agile method.
No more need to wait 10 years of building and at the end SQ tell me it is failed. We just unveil the 10 year box and once we find a strategy cannot pass any one year, we can directly throw it out immediately.
WJPII
3 years ago #268184
Thank you. That was the answer/solution I was hoping for. 3 months of leaving my computer on 24/7 at running 10 year testing, 5 year testing and even 1 year testing is not getting anything close to passing any WF tests. Even after optimizations over and over again. And then there is the problem with the amount of trades.
So right now, I am experimenting with what you just said with one exception (I will explain): I am doing just 6 months in the builder, to get at least 1000 strategies with the simulated 1 tick mode. The exception to yours – and it is a big one – is the TF. I am starting at H1. You hit the nail on the head about the win rate. M1, M5 and even M15 has a much lower win rate in general even though it produces more trades. But in SQX, I am finding that if you do M15 or below for a longer test period (1 year +), SQX is not picking up all signals and hence, you are still getting the same amount of trades as H1, but with a lower win rate. The trade off obviously is the DD when going to H1. But, at this point, I am willing to have a higher DD (10 to 30%) as long as I can get an EA that will have a >80% win rate at the 10 year back test. After the 1000 strategies in builder with the simulated ticks, I am then going to send those to restester for the same time period of 6 months (same date range as the builder for now), but move to real tick back test ONLY. I move the ones with the W/L of >3 to another tab to retest again, but then go to 1 year with real ticks. I then will weed those out with the same win rate % to another tab. At this point, I only have 5-20 strategies to work with. So, now I am going to leave it at the 1 year, real ticks, but use the cross checks of higher back test precision, MC, Opt profile (3000 ops) and WF Op. That is where I am at right now. However, this step is taking about 3 hours per strategy. So I have 22 strategies that I will need to leave me computer running for another 2 days approximately (it has already been 24 hours). My plan then is to take the ones closet to meeting all the tests (none of them ever do actually) and move them to the optimizer for just 1 year @ 5000 passes. I am thinking there will only be 5-10 strategies remaining at this point. Probably all with have a slighter better performance at the end of this, then start the whole process over again for another 1 year (like 2018). Then 2017, then 2016, etc for up to 10 years. It is going to be a painstaking process to do all of this for just 1 EA that will be able to have a decent result for a 10 year back test in MT4. But, it looks like I have no choice. Beause if I try to scale up from 1 year to 5 to 10 in SQX, every single EA fails after optimization AND 2/3 of the trades go away. So I hopefully will be able to get 1-2 EA’s past 10 cycles of testing and optimizing on a 1 year test period in SQX to move to a 10 year back test in MT4 that will have decent results. At least that is what I hope because nothing I have done thus far works in SQX. Granted, I can get an EA out of SQX after optimizing that will back test in MT4 with similar results. But again, it only has 1-2 trades a month. Sometimes it would not trade for 3-4 months at a time. And that is with a lousy 75% win rate (best I have gotten in 3 months!). Also you noted about the computer. I have the latest generation AMD CPU with 64GB ram and all the bells and whistles. $4000 liquid cooled gaming console that I have set up to run 100% of CPU cores and 100% of RAM just for SQX. And it STILL takes months to go through all of these simulations to get absolutely nothing. And then there is the problems with the 1-2 floating trades even though I have everything set up right. It doesn’t happen in the 1-2 year back tests nor in MT4. But, when going out 3 to 10 years, I am always getting massive floaters which then the EA will fail in SQX tests because of overlapping trades.
To be honest, I also have Forex Strategy Builder and it is SO much faster and better at making EA’s that actually trade nicely. BUT – you cannot make EA’s with FSB that will make it past the validation on MQL5 because FSB EA’s takes multiple days to back test due to the 10,000 lines of code it generates.
I am totally pulling my hair out over spending $8000 in hardware and software to make EA’s and cannot get any with good results and into MQL5 to sell….
Thanks again for your input. I am doing all of this right now. See what happens. If this doesn’t work, I guess I will need to go to coding school and give up on EA building software.
Waid
3 years ago #268191
Thank you. That was the answer/solution I was hoping for. 3 months of leaving my computer on 24/7 at running 10 year testing, 5 year testing and even 1 year testing is not getting anything close to passing any WF tests. Even after optimizations over and over again. And then there is the problem with the amount of trades. So right now, I am experimenting with what you just said with one exception (I will explain): I am doing just 6 months in the builder, to get at least 1000 strategies with the simulated 1 tick mode. The exception to yours – and it is a big one – is the TF. I am starting at H1. You hit the nail on the head about the win rate. M1, M5 and even M15 has a much lower win rate in general even though it produces more trades. But in SQX, I am finding that if you do M15 or below for a longer test period (1 year +), SQX is not picking up all signals and hence, you are still getting the same amount of trades as H1, but with a lower win rate. The trade off obviously is the DD when going to H1. But, at this point, I am willing to have a higher DD (10 to 30%) as long as I can get an EA that will have a >80% win rate at the 10 year back test. After the 1000 strategies in builder with the simulated ticks, I am then going to send those to restester for the same time period of 6 months (same date range as the builder for now), but move to real tick back test ONLY. I move the ones with the W/L of >3 to another tab to retest again, but then go to 1 year with real ticks. I then will weed those out with the same win rate % to another tab. At this point, I only have 5-20 strategies to work with. So, now I am going to leave it at the 1 year, real ticks, but use the cross checks of higher back test precision, MC, Opt profile (3000 ops) and WF Op. That is where I am at right now. However, this step is taking about 3 hours per strategy. So I have 22 strategies that I will need to leave me computer running for another 2 days approximately (it has already been 24 hours). My plan then is to take the ones closet to meeting all the tests (none of them ever do actually) and move them to the optimizer for just 1 year @ 5000 passes. I am thinking there will only be 5-10 strategies remaining at this point. Probably all with have a slighter better performance at the end of this, then start the whole process over again for another 1 year (like 2018). Then 2017, then 2016, etc for up to 10 years. It is going to be a painstaking process to do all of this for just 1 EA that will be able to have a decent result for a 10 year back test in MT4. But, it looks like I have no choice. Beause if I try to scale up from 1 year to 5 to 10 in SQX, every single EA fails after optimization AND 2/3 of the trades go away. So I hopefully will be able to get 1-2 EA’s past 10 cycles of testing and optimizing on a 1 year test period in SQX to move to a 10 year back test in MT4 that will have decent results. At least that is what I hope because nothing I have done thus far works in SQX. Granted, I can get an EA out of SQX after optimizing that will back test in MT4 with similar results. But again, it only has 1-2 trades a month. Sometimes it would not trade for 3-4 months at a time. And that is with a lousy 75% win rate (best I have gotten in 3 months!). Also you noted about the computer. I have the latest generation AMD CPU with 64GB ram and all the bells and whistles. $4000 liquid cooled gaming console that I have set up to run 100% of CPU cores and 100% of RAM just for SQX. And it STILL takes months to go through all of these simulations to get absolutely nothing. And then there is the problems with the 1-2 floating trades even though I have everything set up right. It doesn’t happen in the 1-2 year back tests nor in MT4. But, when going out 3 to 10 years, I am always getting massive floaters which then the EA will fail in SQX tests because of overlapping trades. To be honest, I also have Forex Strategy Builder and it is SO much faster and better at making EA’s that actually trade nicely. BUT – you cannot make EA’s with FSB that will make it past the validation on MQL5 because FSB EA’s takes multiple days to back test due to the 10,000 lines of code it generates. I am totally pulling my hair out over spending $8000 in hardware and software to make EA’s and cannot get any with good results and into MQL5 to sell…. Thanks again for your input. I am doing all of this right now. See what happens. If this doesn’t work, I guess I will need to go to coding school and give up on EA building software.
Hi WJPII,
I recommend that you may consider revise your pipeline to the following to achieve the maximum performance since the time it spends will be significant difference:
first, don’t do the whole pipeline manually, since you will need to process the “pipeline loop” at least hundred of times (i.e., you will need to execute the pipeline hundreds of time to find a really workable EA).
So build the pipeline at “custom projects” at SQ’s left panel by the following:
1. Add new task: build strategies, to build 6 months of low TF (My personal thinking is that H1 is too high for selling EA, since you will got very few trades in a month), and catch around 2000 strategies in your databank. (note that you don’t need to set your filter Ret/DD since you are trying to build ‘high SL, low TP’ strategy, so instead, high win% is your core filter), precision= 1min if your strategy TF is set to 5/10/15min
2. Add new task below the prev. one: Retest strategies on another random year within your pre-set 10 year build range. Turn on “Delete failed strategies from databank” in “Ranking” tab. also keep the prevision = 1min, since this step is consider as a “build” step follow by step 1, so don’t change anything (also don’t change filter) like you do in step 1.
3. Add same task as step 2. with different year.,
4. same task, do it until whole ten years are finished.
5. set a threshold for the minimum number for this 10 years “build” strategies before moving on to next move. for instance, pick up ‘Go To Task’ task, and set as: number of result in databank, databank is lower than 80. in this way SQ will automatically re-initiate the whole build pipeline until your databank is filled with 80 strategies.
6. now you can start pre/last OOS, MC tests, WFM test for those passed strategies, even with high precision like you did.
My PC setup is AMD 3950x with 64G ram. 15 cores are running. So after my current build, maybe I will also run one to help you figure out whether this setup work or not.
Waid
3 years ago #268192
Add one correction, from step 1 to 4, the precision should set to ‘select timeframe only’, and turn on cross check with 1min (or even 1 tick)
hankeys
3 years ago #268195
you need to reconsider your approach and simplify the procedures you are using
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.
WJPII
3 years ago #268196
Thank you for this work flow. So much better than the manual approach. I am starting this today. The flow I have is:
1. build: GBPUSD M5 6 month time period 2BE15 TP min 100 max TP, same on SL.
2. retest 1 year random BT
3. retest 1 year random BT (different year)
4. restest 10 year
5. Optimize 10 year
6. retest 10 year multi pair
I can already tell this is going to take weeks of leaving my computer on 24/7 to get to number 6. And even then, I doubt I am going to get any EA’s that will work. I am foregoing and WT or MC just to see if I can get anything our of this process. It is nice to have this custom project function, but what a mess this is just to produce a workable EA. I would be shocked if anyone who has purchased this software actually created anything that is usable for both sale and in a live account.
And here is what is sad….I have gone through charts for weeks to find a combo of indicators that have produced very good entry/exit points and can trade manually with that with a 95% win rate. But, can’t automate the same thing with this software.
Thanks again. Maybe I will get something in a month from now or smash my computer…
WJPII
3 years ago #268197
Thank you again for your input. I have tried everything to get anything out of this software that would have a 1+ profit score over a 10 year BT. It doesn’t matter what I do – it has been impossible. Yes – I could maybe get something that has a trade 1 or 2 times a year – but who trades like that??
If you know of a better work flow – please let me know. Would love to hear it.
WJPII
3 years ago #268215
Well, the flow we discussed is not working and not sure why. Has to be in one of the settings, but I cannot figure out which one. Basically, after 6 month Building strategies, I send them to optimization for 1 year before sending them to retesting for the same year. Well, the EA’s are not even getting to the retesting part of the flow. They are just stuck in the optimizing (only 200 times each x 50) on selected time frame only for ticks. Each optimization is only taking a couple of minutes, but this has been going on for 3 hours. They are not moving to the next folder for restesting. This sucks. Any suggestion? See screenshots.
WJPII
3 years ago #268216
And by the way – when I did the build at 6 months and tried to go directly to restesting at 1 year – all failed even though the only test to pass was a profit score of 1. They go from builder at 6 months (1000 of them) with profit score of 1.5 to 4 to <1 after 1 year retest. So I had to add optimization. Which now they are not leaving that part of the flow.
Waid
3 years ago #268457
Hi WJPII,
By the recent try errors, I have concluded that the settings of ‘build config’ and ATM plays an very important rule.
Different currency pairs will definitely set to be different according to the ATR value, esp. when you have fixed SL or TP. Larger ATR pairs like GBPJPY need large value to achieve the building speed. It will accelerate more than 50% than the original if you set appropriate values.
Also if you think you are always knocked out by robustness test, try to adjust ATM ratios/SL. But I think ATM is more appropriate to be put into exit rules in building blocks tab since it is just another variance of exit.
Good ATM settings will have more smooth equity curve than the original one for sure. And the equity curve will strongly impact the robustness test especially for MC tests.
Viewing 14 replies - 1 through 14 (of 14 total)