GSB Forums - Powered by XMB 1.9.11

At long last, I publish my default settings I am using for testing.
All files including data files are included, but not all the gsb exe,dll, xml files.
Included is gsb.config.xml to get GSB into admin mode.

There may be some settings that need a tweak, so consider this the first revision.
There are 4 ES files with 5 ticks of noise, 4 ES files with 10 ticks of noise

I think you can put this into the folder of your choice. (All paths relative, not fixed)
200mb in size zipped up.

My top 11 indicators are selected, and these are better than any other method Ive found. However sorry to say, im not sure how I found this group.

All macros used are included.

In summary you build 50,000 systems. Dates are 2000 to 2015.6.30 nth 1 / 80
The manager then populates but with nth set to all (workers built them with nth no trade)
Then macro 1 selects the top 250 systems and walk forwards them.

Manually put the top 250 into favouritesD.
Then run macro3 to get the stats.
record these stats.

Then run the macro5 astab-c macro that deselects all favouritesD, and puts the top 90 astab-c
Then run macro3 to get the stats.
The record these stats.

go to favourites A
Then run macro2 which changes dates back to 2015.6.30
after this is complete, run macro6 which selects the top 91 vss systems of the top 250
Then run macro3 to get the stats
Shown below is a summary of my stats.

zip file here of all my files in private forum
https://trademaid.info/forum/viewthread.php?tid=241#pid4133

Hi Peter,

I used your settings and followed your steps in building ES system.s I see some good result on the statistics, I am not sure if I cheated or just lucky.

The systems built seems to have a low fitness number.
I have grouped the top 91 systems into families and the family with highest number of members has only 5 members.

I have attached some screenshots, let me know how you interpret this result.

Thanks

admin - 12-9-2019 at 04:14 PM

Appengineer,
those numbers are higher than I could achieve, and I have spend massive time on these tests. 16,000 is about as high as I get, and thats rare.
You got a total of 21270. The most obvious human error is you did a wf from 2000 to 2018, not 2000 to 2015.
Im happy to look at teamviewer to confirm what your doing. The other alternative is you dodn something brilliant.
Another option is you data ends now, not feb 28 2019 (which mine does)
There have been some decent wins since feb 2019

appengineer - 12-9-2019 at 04:43 PM

Appengineer,
those numbers are higher than I could achieve, and I have spend massive time on these tests. 16,000 is about as high as I get, and thats rare.
You got a total of 21270. The most obvious human error is you did a wf from 2000 to 2018, not 2000 to 2015.
Im happy to look at teamviewer to confirm what your doing. The other alternative is you dodn something brilliant.
Another option is you data ends now, not feb 28 2019 (which mine does)
There have been some decent wins since feb 2019

I will send you team viewer details. Something is off, the pearson(R-F) is also too low.

admin - 12-9-2019 at 04:49 PM

appengineer, the metrics you show are from the 2018 to 2019 period, thats why the metrics are so low. To me the bigger issue is why are your results so (too??) good
You cant get high pearsons over a really short period. Thats not a issue at all, just a mathematical fact

appengineer - 14-9-2019 at 09:01 AM

appengineer, the metrics you show are from the 2018 to 2019 period, thats why the metrics are so low. To me the bigger issue is why are your results so (too??) good
You cant get high pearsons over a really short period. Thats not a issue at all, just a mathematical fact

Thanks Peter, I have started to understand the process a little better.

Can you share the best stats you have achieved so far for ES? This will provide an idea of what we can aim for.

If the Trades Count is low (Full period max=93) for the final systems chosen after completing all the steps. Is the any need to evaluate the top system in each family or you just start a new build?

admin - 15-9-2019 at 07:57 PM

what youve done is very good, and I dont get much better than what youve done. Occasionally you get a result thats a moderate amount better, but then its not repeatable.

One comment. There are new bugs in MC, and some bugs in GSB. The any indicators cross is faulty. Not many users have used this. Much of this will be fixed in the next build

appengineer - 17-9-2019 at 08:26 PM

Just to share, I think I'm onto something, here are the result from today's build for ES.

Peter - once we import the strategy on TS, what verification do you perform to make sure GSB and TS results are the same?

admin - 17-9-2019 at 08:38 PM

appenginer, the metrics in the ts code should match whats on ts. 98%+ the same is ok.

appengineer - 19-9-2019 at 07:37 PM

Looks like a good ES system OOS from 03/01/2018 to date, what do you say Peter?

Attachment: Login to view the details

admin - 19-9-2019 at 08:25 PM

Looks like a good ES system OOS from 03/01/2018 to date, what do you say Peter?

It looks good from what I can see. Same observation on my systems is that most profit in last few years is short.

appengineer - 21-9-2019 at 09:57 PM

Peter, These latest updates are working great!

I just wanted to share a really simple build that is delivering solid results. Here I've just used the ES30 with the NQ30 and YM30 (instead of ES29,ES31), followed the same process you've outlined in the help using your Macros and the results can be seen. I know you are likely doing better, however, I'm just a hack and let's not forget you're the pro! :cool:

Also seeing positive results when using ES30 with NQ30 and YM30. Verification Data has random noise on ES30

admin - 22-9-2019 at 09:05 PM

I dont feel like a pro, been chasing my tail for a long time trying to dig down very deep to how things are working, but feel I made progress today.
I would like to know what you did differently here.
Just ym nq in replacement of es29 and es 31?
Your c vs d results are much higher, likely because the wf is only done on es 30
(thats fine but I find what youve done interesting)

Was it just the verification included extra markets, or was it the random noise and the extra markets?

Had a thought provoking weekend. Assumptions being challenged (in areas not to do with trading) Chickens, termites, dogs & talking parrot. In life there a little clues we ignore.
In Vanuatu meet a very enterprising educated guy. He says the chicken he wants to kill to eat, is normally in the yard further-est away from him, hiding behind all the other chickens. This a few 100 meters away from where the survivor show was filmed.
In the past I would have ignored the fact that a specific chicken (are really dumb) would know sometome wants to kill them.. My regret however is I found out too late he owns a shot gun and 22 rifle. We stayed in an area with lots of illegal roosters that decided to sing at 4am. Have never wanted a gun so badly. Roosters illegal as its a residential area.

Also seeing positive results when using ES30 with NQ30 and YM30. Verification Data has random noise on ES30

appengineer - 24-9-2019 at 07:17 PM

Also seeing positive results when using ES30 with NQ30 and YM30. Verification Data has random noise on ES30

Hi Peter,

Not sure if the response is directed to me but
1. Yes,I only replaced es29 and es 31 with ym nq
2. Verification was only on ES with random noise, NO extra markets.

Thanks

admin - 24-9-2019 at 07:31 PM

Hi AppEngineer,
yes it was redirected to you. Your results are better than the final figures show ie total CDEF 11438 as the final figures aveD 4107.. and AveF 3360 are extremely high and represent a walk forward done only on ES, but out of sample. Im going to do more testing on this down the track. Well done.

appengineer - 24-9-2019 at 08:05 PM

Here is another okay ES system, seems to be going flat lately though

Attachment: Login to view the details

admin - 27-9-2019 at 01:47 AM

Here is another okay ES system, seems to be going flat lately though

Keep up the good work

Daniel UK1 - 27-9-2019 at 02:04 AM

@appengineer and Peter, after some tests i find that systems is much harder to build using NQ30 and YM30 instead of ES 29, 31... Is this your finding also? not saing its not good, i just mean that it takes longer time to build.. thanks for sharing

admin - 27-9-2019 at 04:49 PM

@appengineer and Peter, after some tests i find that systems is much harder to build using NQ30 and YM30 instead of ES 29, 31... Is this your finding also? not saing its not good, i just mean that it takes longer time to build.. thanks for sharing

Yes, and that is no surprise. Its harder to find a system that works on other markets without degradation. You can loosen the pearsons and or pf figure a little.
Im testing this method right now, and will let you know the results

admin - 28-9-2019 at 12:35 AM

Also seeing positive results when using ES30 with NQ30 and YM30. Verification Data has random noise on ES30

Did you try other combinations? rty(Russell 2000) and emd was not as good.
My results were similar to yours

appengineer - 28-9-2019 at 08:56 AM

Hi Daniel,

Yes it takes a while to build but I normally leave it to run over night.

Thanks

appengineer - 28-9-2019 at 08:59 AM

Also seeing positive results when using ES30 with NQ30 and YM30. Verification Data has random noise on ES30

Did you try other combinations? rty(Russell 2000) and emd was not as good.
My results were similar to yours

Thanks Peter - I haven't tried another combination yet, still running builds using NQ and YM to see if the performance is consistent, so far that's the case.

Good to know RTY and EMD are not as good.

admin - 1-10-2019 at 01:57 AM

I tried YM, nq, emd rty (ER2000), but not exhausive combination
The best symbol by itself was NQ, with a more distanct YM
NQ YM with ES however were looking like they had some promise. However Im not decided either way on this issue yet

appengineer - 14-10-2019 at 07:55 PM

Hi Peter,

Could you explain the nth day concept?

My understanding is GSB will take a trade every 80th day if there is a signal generated on that day, is that correct?

If so, the training data becomes so small given there is only 253 trading days a year.

Thank you.

admin - 14-10-2019 at 08:00 PM

Hi Peter,

Could you explain the nth day concept?

My understanding is GSB will take a trade every 80th day if there is a signal generated on that day, is that correct?

If so, the training data becomes so small given there is only 253 trading days a year.

Thank you.

Its going to average out to 50% of the days in the years chosen are out of sample. Generally the various people who know what they are doing and promote their own methodology often use high out of sample figures than this. GSB however has a lot more validation methods than most other people.

Bruce - 15-10-2019 at 03:25 AM

This is all really good feedback, I tried related markets, (across metals, bonds & T-notes, indexes) just couldn't get anything to deliver a better result than the well documented 29,30 and 31min.
I've recently taken a slightly different tack using a similar methodology to Peters however building on more recent data frame (2010-12312017 IS, 01012018-06312019 OOS) and only focusing on the fittest of systems on OOS data only over the last 18 months or so.

I'm only interested in the OOS output, in that I'm building 50k systems and have the App optimization setting for AutoDate Mode set to NoTrd. Then I use a macro to filter the top 5000 x fitness, change Date Mode to all, filter the top 1000 for WF, then filter on WF OOS to get top systems. Expose the best systems to totally unseen data and then selecting the best performers be that NP/DD, Pearsons etc. to PA for final periodical selection.

It's working ok, early days with this, however NQ is working really well, building TPD 30min and 3 x random 5 ticks and 3 x random 10 ticks (for NQ).

Just a slightly different spin, have used this method to build a trade portfolio with, ES, YM, US, TY, GC & HG.

saycem - 16-10-2019 at 11:53 PM

I finally got a chance to try this. My results attached. Seem lower than others?

admin - 17-10-2019 at 12:20 AM

Hi Saycem,
my published results were 11438, and yours are 13133 so I think what you have is fine.
Have I miss-interpreted this?

Peter

saycem - 17-10-2019 at 01:03 AM

Thanks yes I misread yours.
Scrolling through some of the results in Fav D after running macro 6 (VSS) and then macro 3 (stats) and changed Dates to All.
I'm seeing two different types of curves (attached). One has WF to 2015 other is WF 2018-2019? Looks like an error maybe?

admin - 17-10-2019 at 01:35 AM

Thanks yes I misread yours.
Scrolling through some of the results in Fav D after running macro 6 (VSS) and then macro 3 (stats) and changed Dates to All.
I'm seeing two different types of curves (attached). One has WF to 2015 other is WF 2018-2019? Looks like an error maybe?

No Likely if oos over ride org settings is on, and use wf params is on, the wf curve is extended using the parameters in the last run of wf. (2015)

saycem - 17-10-2019 at 09:28 PM

I realise there may be no right answer to this, but I'm interested in what groups to look at to select systems for live trading.

Do you choose from top 2 Family of Fav D after Astab (macro 5)?
or from top two family of Fav D after VSS (macro 6)?
Or do you look at whole 250 WF and pick something with high VS and Astab?

admin - 17-10-2019 at 09:48 PM

I realise there may be no right answer to this, but I'm interested in what groups to look at to select systems for live trading.

Do you choose from top 2 Family of Fav D after Astab (macro 5)?
or from top two family of Fav D after VSS (macro 6)?
Or do you look at whole 250 WF and pick something with high VS and Astab?

Thats a good question that I have not yet put so much work into.
I would choose the families that have a reasonable amount of members in it.
say you have 250 systems,
the top family might have 30 systems. So system1 of family 1 is my top pick, though other members in the family lower down MIGHT be better.
It would be valid to choose from the 250 or the 90 (top astab) or the 91 top vss.
My thoughts are maybe to choose the one that gave the best metrics out of
the top 250, 90 (astab), 91 (vss)
so lets say top 250 total is 12,000, astab(90) is 12800, and vss (91) is 12300
I would choose systems / families from the top astab 90

saycem - 20-10-2019 at 08:05 PM

Can I confirm that the systems saved in Favourites B and C are not used for anything in this process?

admin - 20-10-2019 at 08:15 PM

Can I confirm that the systems saved in Favourites B and C are not used for anything in this process?

Not unless you want a wider , larger rangeof systems

saycem - 20-10-2019 at 08:52 PM

Thanks - a couple of questions on system selection.

How important do you think it is to pick systems without looking at performance from 2016-2019?

If you choose not to look at 2016-2019 performance when selecting once you've picked a system would you then WF all the way through to 2019 before trading live?

admin - 20-10-2019 at 08:57 PM

Thanks - a couple of questions on system selection.

How important do you think it is to pick systems without looking at performance from 2016-2019?

If you choose not to look at 2016-2019 performance when selecting once you've picked a system would you then WF all the way through to 2019 before trading live?

I think I would like to look at all performance, even from 2/2019 to todays date.
I likely would do the 2/2019 till todays date at the end of the process as an extra check and balance
IM only doing wf from 2000 to 2015.6 (unless i do a complete wf back in TS (and i dont do that often)
ne of the reasons is , 2018 -2019 was such a wild year, systems often want a parameter change for those years. I dont like the idea of parameter change for just one year. In late 2008 we had the same issue due to extreme profits in 2007-mid 2008

saycem - 20-10-2019 at 09:26 PM

Thanks Peter that's interesting about 2018 and makes sense.
So when picking a system you change Global Dates to ALL and look at the original system to 2019 and the WF to 2016. How do you see beyond 2/2019 in GSB when there is no price data in the price file. It would take to long to load every individual system into TS no?

I just want to distinguish between checking the post 2016 vs using it to determine what system to trade. I'm worried if once you compare oos data to pick a system it's no longer true oos if you use it to cherry pick the best past performance?

admin - 21-10-2019 at 11:19 PM

Thanks Peter that's interesting about 2018 and makes sense.
So when picking a system you change Global Dates to ALL and look at the original system to 2019 and the WF to 2016. How do you see beyond 2/2019 in GSB when there is no price data in the price file. It would take to long to load every individual system into TS no?

I just want to distinguish between checking the post 2016 vs using it to determine what system to trade. I'm worried if once you compare oos data to pick a system it's no longer true oos if you use it to cherry pick the best past performance?

Note quite
I use the wf settings to 2019, but the actual wf was only done to 2015
You can update your price files, and do override original settings, but Im just putting the system into TS. Yes the OOS issue is always with us humans.
The easier it is too look at oos, the more likely we curve fit. Hence I make things hard for myself and are using TS

admin - 21-10-2019 at 11:21 PM

The next video is due out in < 48 hours. It covers all the entry types and weights & various issues. The video is so over due to be produced, and there has never been a video thats taken as long to produce as this one, and the few that will follow I hope in the near future. Months of work & research.

admin - 23-10-2019 at 01:19 AM

the next video is out. Thanks for the thanks

https://www.facebook.com/Genetic-System-Builder-by-Trademaid...
comments on fb and forum welcome.

Next will be the market validation video. Basically a video on whats here

There are also small but important to read updates here
https://trademaid.info/gsbhelp/Updates.html

Daniel UK1 - 23-10-2019 at 03:07 PM

Thanks for the Video Peter, so looking at the recommended settings etc according to the methodology, it would be interesting to understand if this is the recommended settings according to you for ALL symbols or just ES?

admin - 23-10-2019 at 03:32 PM

Thanks for the Video Peter, so looking at the recommended settings etc according to the methodology, it would be interesting to understand if this is the recommended settings according to you for ALL symbols or just ES?

That's a really good question, that Im keen to know. I dont have resources to tell you that now, but your welcome to try yourself. IM working on the next video now.

My notes from the 26 Nov 2019 Methodology Video

JasonT - 2-12-2019 at 02:18 AM

Hi folks,

I've taken notes from Peter's most recent methodology video. It wasn't much more effort to convert them into a little guide. Peter has been kind enough to go through them to confirm the steps are as intended.

I'm posting them here in the event that others may find them useful. Please be aware they are simply notes and not detailed step by step instructions. That's what the video is for.

Happy for your any feedback or suggestions for improvements.

Attachment: Login to view the details

admin - 2-12-2019 at 09:35 PM

Here are the macro's used in the video.
Please confirm they work ok.
unzip and copy to C:\GSB\Data\Settings\Macros
m1 build systems, wf top 250
m2 change dates to 20150630
m3 do stats: 20190228 is added as the end date for stats. (this is to avoid accidentally changing the benchmark from end date 20190228 if you update data.)
m4. None
m5. Put top 90 astab-c in Fav D
m6. Put top 91 vss in Fav D

Attachment: Login to view the details

Live traded SI system built on GSB

Daniel UK1 - 4-12-2019 at 08:01 AM

Sharing results from a live traded system built on GSB if its of interest, very robust. Built for QSI Silver futures for 30 min data. Built using data only from 2007 to 2015 /02/28... validates great on QSI 60, 120 and 240, and on QCL 60 and 30 etc... First pic is traded system, last picture is stats from real traded system and rest is validation on different timefrarmes and markets such as oil futures etc.

admin - 4-12-2019 at 04:46 PM

Sharing results from a live traded system built on GSB if its of interest, very robust. Built for QSI Silver futures for 30 min data. Built using data only from 2007 to 2015 /02/28... validates great on QSI 60, 120 and 240, and on QCL 60 and 30 etc... First pic is traded system, last picture is stats from real traded system and rest is validation on different timefrarmes and markets such as oil futures etc.

Great to see such postings. Have you tried the new methodology on silver now, and compared to how things were built in the past?

Daniel UK1 - 5-12-2019 at 03:08 AM

Great to see such postings. Have you tried the new methodology on silver now, and compared to how things were built in the past?

This systems was built in the early days with me using GSB, however still using strict validation and picked from systems with over 70 PAS that i was using to narrow down to final pick group. so no, QSI needs to be re tested using latest methodology to see if better results can be achieved.

admin - 5-12-2019 at 08:16 PM

with the macros & methadolgy, my data ended 20190228
However other users are going to have later dates, (not necessarily deliberately)
If thats the case, to keep the benchmark the same for all users, and to give another period out of sample for us to check our work,
macro3 needs a date change.
The end result was the user did a really good job of testing, but tested a few months later

RandyT - 18-12-2019 at 03:58 PM

Just noticing that my M3 macro is a bit different. Is there a place to grab current versions of these? Not sure what ParamStats does and if all of the Idle steps are necessary in my version of the macro.

RandyT - 18-12-2019 at 04:02 PM

Working on my end to get through one pass of this new methodology. Could someone provide more detail as to where the numbers come from that are shown in this spreadsheet? I've watched every video I can find and don't really see this specific location or process to get these numbers.

I also am not seeing differences in the values in the Statistics tab after M3 macro runs which makes me suspicious of the current version of the macro I am using.

Thanks

admin - 18-12-2019 at 04:53 PM

Working on my end to get through one pass of this new methodology. Could someone provide more detail as to where the numbers come from that are shown in this spreadsheet? I've watched every video I can find and don't really see this specific location or process to get these numbers.

I also am not seeing differences in the values in the Statistics tab after M3 macro runs which makes me suspicious of the current version of the macro I am using.

Thanks

I'm hot air ballooning now so can't help till later today. The November 2019 methadolgy should have it and the docs on trademaid.info. Look also under legacy docs

RandyT - 18-12-2019 at 05:03 PM

I'm working from this doc. https://trademaid.info/gsbhelp/Methodology.html

I've looked at some of the legacy docs there and do not see an answer to my question. I will need a bit more help to get past this.

FWIW, I've taken a copy of the Methodology doc, have created a Google doc and am making changes to this as I work through issues to add more detail, fix typos and hopefully make it more clear for the next noob.

https://docs.google.com/document/d/1BEtZNssuq_FpRL-SqqxVPDCn...

If anyone would like edit permissions for that doc, just let me know.

admin - 18-12-2019 at 10:43 PM

Here is an example of the stats

RandyT - 18-12-2019 at 11:04 PM

Ok, so the numbers shown in the spreadsheet image in the methodology doc are * 1000 then. Could not figure out where they came from.

In far right cell, it says "Total D+E+Ave", but seems it is the total of Avg. C, D, E and F. Correct?

admin - 18-12-2019 at 11:36 PM

I'm working from this doc. https://trademaid.info/gsbhelp/Methodology.html

I've looked at some of the legacy docs there and do not see an answer to my question. I will need a bit more help to get past this.

FWIW, I've taken a copy of the Methodology doc, have created a Google doc and am making changes to this as I work through issues to add more detail, fix typos and hopefully make it more clear for the next noob.

https://docs.google.com/document/d/1BEtZNssuq_FpRL-SqqxVPDCn...

If anyone would like edit permissions for that doc, just let me know.

Hi Randy
correct on all points. I have updated all the doc
I updated the docs to answer your questions. You need to refreash the browser.

RandyT - 19-12-2019 at 12:09 PM

I can also confirm that one of the problems I was having with getting correct results was a problem with the macros I was using. I've grabbed a copy of these from a clean installation and have replaced all macros with versions that are prefixed with macro number. The confirmed problem was in the M3 macro.

Not sure if others were contributing to the result problems.
Not sure if these are the latest.
Might be a good idea to consider versioning this somehow.

RandyT - 19-12-2019 at 05:58 PM

I was doing some indicator test runs and discovered that in the in the optimization settings file that was supplied with GSB install (ES-IndicatorTesting.gsboptset) you are setting 3 entry modes. In the latest video on the indicator setting topic that starts at 9:50 in the video you are using one entry mode.

Which is the current best practice?

With the 3 entry modes I get the indicators shown in first image.
With 1 entry mode, I get indicators shown in second image.

admin - 19-12-2019 at 07:46 PM

I was doing some indicator test runs and discovered that in the in the optimization settings file that was supplied with GSB install (ES-IndicatorTesting.gsboptset) you are setting 3 entry modes. In the latest video on the indicator setting topic that starts at 9:50 in the video you are using one entry mode.

Which is the current best practice?

With the 3 entry modes I get the indicators shown in first image.
With 1 entry mode, I get indicators shown in second image.

I went through your list and see there is two indicators swapped in the two lists. I looked up my old config file and saw that I used crossdualEntryLevels only on ES. One thing to note is that the identical test may give some variation in 1 or 2 indicators (the weaker ones)

However seeing as CL (crude oil) likes compare1 & compare2 it might pay in the methodology, to determine the best entry type first. This will only need to be done first time we do tests on a market.

admin - 19-12-2019 at 10:40 PM

I can also confirm that one of the problems I was having with getting correct results was a problem with the macros I was using. I've grabbed a copy of these from a clean installation and have replaced all macros with versions that are prefixed with macro number. The confirmed problem was in the M3 macro.

Not sure if others were contributing to the result problems.
Not sure if these are the latest.
Might be a good idea to consider versioning this somehow.

When Im back from Vietnam (Christmas time) I will make a separate thread that has the most current macros.
This will get much more useful after settings-full macro is done.

Daniel UK1 - 27-12-2019 at 03:37 AM

Optimised entry... in MC we have a setting that can use limit orders for n seconds and then convert them to market if not filled.... have anyone using Multicharts, done some tests to see if using limit order and then convert those to market if not filled within n seconds, has any positive impact on fills for live traded systems ? i would be interested to hear if anyone have done some tests...

admin - 27-12-2019 at 03:57 AM

Optimised entry... in MC we have a setting that can use limit orders for n seconds and then convert them to market if not filled.... have anyone using Multicharts, done some tests to see if using limit order and then convert those to market if not filled within n seconds, has any positive impact on fills for live traded systems ? i would be interested to hear if anyone have done some tests...

TS has the same feature. I dont trade with TS, so cant comment.
Im skeptical if it helps much on entries, but think its likely it helps at moc time.
Moc is volatile,and aiming for 1 or 2 ticks more I feel is a good idea.

RandyT - 7-1-2020 at 02:02 PM

Peter,

Wanted to followup this post to ask if you could give a little help as to how to go about testing for best entry mode? Would (could?) this be done with no indicators selected and all entry modes selected? Is there a similar way of looking at the statistics on entry modes as there are parameter performance?

tks

Tutoring on results requested

RandyT - 7-1-2020 at 02:13 PM

Peter,

I was hoping you could take a minute to look at some of the output from a recent run to give me some idea of what I may have done wrong to get the following results. I've seen this on a couple of occasions and suspect I have done something wrong. I'm attaching the opt settings I have used and a couple of screenshots.

1. Verification seems to have failed miserably. Not sure if that just happens on occasion or if I have something wrong in setup. As you can see, the verification score is 0/8 for every system.

2. In attachment showing the statistics for each of the different statistic runs you outline in your new methodology, can you solidify for me what you would focus on given the output? (broad assumption it is valid)

3. Would be very open to any pointers anyone has about the GC market.

Thanks

Attachment: Login to view the details

Daniel UK1 - 7-1-2020 at 03:34 PM

Hi, i can just share my opinion about my own testing for GC, very very difficult and i have not managed to get any good results at all.. very very high degradation, bad validation numbers, and bad NP/DD..
however SI with some aid of GC has been my route to something tradable with pretty good numbers... not the other way around..

I am sure though other have done much better and most likely managed to get GC to work..

Just my 2 cents

I someone managed to get GC work, please share some pointers

RandyT - 7-1-2020 at 03:48 PM

Hi, i can just share my opinion about my own testing for GC, very very difficult and i have not managed to get any good results at all.. very very high degradation, bad validation numbers, and bad NP/DD..
however SI with some aid of GC has been my route to something tradable with pretty good numbers... not the other way around..

I am sure though other have done much better and most likely managed to get GC to work..

Just my 2 cents

I someone managed to get GC work, please share some pointers

Interesting feedback Daniel, I was just running a GC/SI/HG trial and noticed that SI was showing at least double the NP of GC and HG. So I am about to give a try with SI as the primary. First running an indicator optimization to see what is working on SI.

admin - 7-1-2020 at 04:06 PM

Hi Randy
I havnt done gold for a very long time, and I didnt have the tools in GSB we have now at the time.
First thing
Try your benchmark without the 390 bar (daily) data
then try with SI HG as data2 (30 min)
Whats the session time used in gold?

RandyT - 7-1-2020 at 04:24 PM

Hi Randy
I havnt done gold for a very long time, and I didnt have the tools in GSB we have now at the time.
First thing
Try your benchmark without the 390 bar (daily) data
then try with SI HG as data2 (30 min)
Whats the session time used in gold?

So, based on your response, it really could be possible to see validation results that poor and still have proper configuration?

I will give GC another try as you suggest. Currently chasing down the path that Daniel mentions regarding SI.

This particular run I am showing is a session of 0330-1330 exchange time. (Had some other research that suggested that) I've also been exploring in the range of 700-1330 which is what I come up with when using your approach to identify higher volume times.

admin - 7-1-2020 at 04:45 PM

Hi Randy
"So, based on your response, it really could be possible to see validation results that poor and still have proper configuration?" Yes thats possible. Depends on the market.
I would test both session times. Sometimes both can be valid, but one better. I like the short session as its more liquid.

Conquering SI market

RandyT - 16-1-2020 at 04:45 PM

Daniel, thanks for this info. I wanted to go into a bit more detail about what I have tried on SI since reading your comments and see if you have any other suggestions as to how to find success with SI.

First, I have been trying on SI market, LongShort, Session 0700-1330 Exchange with 2 indicators and have limited the built-in indicators to the following:

AverageFc
DeCyclerOscillator
DmiMinus
HighestFc
Hurst
KeltnerLowerBand
KeltnerUpperBand
LowestFc
SlowK
StdDev
SuperSmoother

Entry modes seem to only be Compare1 and Compare2.
I have been trying all Normalization modes.
I have experimented with and without StopLoss.

I have been building off of 29, 30 and 31 minute SI data.

Most recently I have been limiting the market timeframe that I am building on from 2012-01-01 to 2018-12-31.

All equity curves seem to degrade over past couple of years. Not clear if that is market or system. I am posting some info to show what I have achieved to see how this measures up. Might be tradeable... not sure.

You say "with some help from GC". I would love to know what you mean exactly. I've tried earlier in this process to pair SI.30 and GC.30 to see what happened but did not seem to improve outcomes.

I'd be happy to share the optimization settings file but seems these refuse to load if data path is not found.

Your help is much appreciated.

admin - 16-1-2020 at 06:39 PM

Hi Randy
here are my two silver systems and bruces silver
I trade live my two, and bruce trades his silver system
mine oos 20190321 but built 2019 04 02
Bruces is a bit older.
all are build on 30 min silver only, but one has 30 min gold data2
zero slippage used.

RandyT - 16-1-2020 at 09:18 PM

Thanks Peter, can you give any hints on entry type or comment on my indicator list?

admin - 16-1-2020 at 10:23 PM

Thanks Peter, can you give any hints on entry type or comment on my indicator list?

I dont remember how I made them, but there could be some postings by user Bruce as he published some stuff months back.
Entrytpye1 was used, but try 1&2, 3 inidicators.

Im going to have to try and find my notes on silver

Daniel UK1 - 17-1-2020 at 02:55 AM

Hi Randy, i think silver is a very good market, however difficult for GSB, but systems seems quite stable. I believe choosing build date is very important since the market itself has been quite different over the data period.. i dont think i was using data after 2015 to build on though.... the steep raise in the equity curve in the middle and the relative to that lowering NP over the past years is because of market and not poor system (in my humble opinion). So i am feeling comfortable trading my SI systems despite this going ahead. With help from GC, i mean that you could use GC as data 2, to help... my research for SI, shows to go for higher timeframes, rather than lower... ... test in that direction and see what you find... my biggest issue with SI is that most of my good systems behaves very similar, and have very high correlation. On the positive side its one of the smoothest systems i have,.... also i have found that it needs rather large stops, min 1000 and up...

Daniel UK1 - 21-1-2020 at 09:46 AM

Hey everyone, just wanted to check if anyone have had any good results on other markets than S, GC, SI, NG, CL and ES the usual suspects... ?
For example currency futures ? or other interesting tradable markets?

admin - 21-1-2020 at 03:38 PM

Hey everyone, just wanted to check if anyone have had any good results on other markets than S, GC, SI, NG, CL and ES the usual suspects... ?
For example currency futures ? or other interesting tradable markets?

Here is xlv long only, inc my fills.
It was not made by me, and a different approach. The author claims GSB works long only on most stocks.
oos 2017 12 27
Please read the forum disclaimer.

rws - 21-1-2020 at 04:37 PM

Hi Peter,

When optimizing on an ETF of a sector, I see far worse results
when walkforwarding than optiziming on all the stocks in that sector in Amibroker. For example with a tribes or particle swarm optimizer. I suspect it would be the same for GSB.

It is very logical, on 1 ticker you are very easily overdefined while
optimizing on 50 tickers, you find more robust OOS parameters settings because you cannot easily overdefine as there is easily a
50-fold more data. You can also have more parameters when optimizing on an portfolio without the risk of curvefitting which can improve the result. You will find less peaks for the parameters right from the start which can be a shortcut instead of many confirmations afterward.

I understand GSB has Data2 but that is not like a portfolio.
I understand GSB can verify on different tickers but optimizing on
different tickers right from the start is usefull.

Would you ever consider adding optimizing on a portfolio in GSB?

Thanks

admin - 21-1-2020 at 06:40 PM

Quote: Originally posted by rws

Hi Peter,

When optimizing on an ETF of a sector, I see far worse results
when walkforwarding than optiziming on all the stocks in that sector in Amibroker. For example with a tribes or particle swarm optimizer. I suspect it would be the same for GSB.

It is very logical, on 1 ticker you are very easily overdefined while
optimizing on 50 tickers, you find more robust OOS parameters settings because you cannot easily overdefine as there is easily a
50-fold more data. You can also have more parameters when optimizing on an portfolio without the risk of curvefitting which can improve the result. You will find less peaks for the parameters right from the start which can be a shortcut instead of many confirmations afterward.

I understand GSB has Data2 but that is not like a portfolio.
I understand GSB can verify on different tickers but optimizing on
different tickers right from the start is usefull.

Would you ever consider adding optimizing on a portfolio in GSB?

Thanks

I agree, and this can be done already.
Build with say your 50 stocks

then wf data, have all 50 datas in there.

You can also do the stats on the entire group of systems too.

Daniel UK1 - 22-1-2020 at 02:33 AM

Here is xlv long only, inc my fills.
It was not made by me, and a different approach. The author claims GSB works long only on most stocks.
oos 2017 12 27
Please read the forum disclaimer.

Hi Peter,

XLV, as in health care fund etf?
Would you be able to share approach direction or how its traded ? for example swing ? intraday?

Thanks for sharing

admin - 22-1-2020 at 03:33 AM

Here is xlv long only, inc my fills.
It was not made by me, and a different approach. The author claims GSB works long only on most stocks.
oos 2017 12 27
Please read the forum disclaimer.

Hi Peter,

XLV, as in health care fund etf?
Would you be able to share approach direction or how its traded ? for example swing ? intraday?

Thanks for sharing

Yes, its swing, health care, and long only and sf was lowest low (unusual)

RandyT - 22-1-2020 at 07:17 AM

Market bias?

admin - 22-1-2020 at 07:23 PM

A minor thing.
If you build your 50k systems, then run macro3 (stats)
you need to run macro2 (dates to 2015) before you run marco5 (top 90 wf) or macro6(top 91 vss)
I only just realized this.

Im going to merge macro2 into the start of m3 to fix this. It will slow m3 down a little.

The reason is m5 and m6 have pf and number of trade filters as well as wf & vss.
The pf filters etc should be done on pre 2015, not post 2015

RandyT - 24-1-2020 at 08:38 AM

A minor thing.
If you build your 50k systems, then run macro3 (stats)
you need to run macro2 (dates to 2015) before you run marco5 (top 90 wf) or macro6(top 91 vss)
I only just realized this.

Im going to merge macro2 into the start of m3 to fix this. It will slow m3 down a little.

The reason is m5 and m6 have pf and number of trade filters as well as wf & vss.
The pf filters etc should be done on pre 2015, not post 2015

This would likely explain why I have been scratching my head regarding the results I am getting with this approach.

admin - 24-1-2020 at 07:48 PM

This would likely explain why I have been scratching my head regarding the results I am getting with this approach.

I only just realized myself. A clue was you didnt get 90 systems into fav D at times.

I didnt think though enough the reason why

admin - 30-1-2020 at 01:16 AM

Don't not use the public cloud for this unless we are on build 56.71 or above.It will produce a error that will stop all cloud workers till they update to 56.71

This will only work in beta tester mode too. (for now)

I have made significant progress on CL this week.
1) Indicator choice
build 30,000 systems. training pf 1.3 pearsons 0.9.
do indicator stats on all 30,000
15 min data with 30 min data2, 14 min data with 28 min data2, 16 min data with 32 min data2,

2) secondary filters
There is a new SF in the alpha build list that for CL is significantly better than closeD.
To find it.
use indicators from 1 above
use 15 min data1, 30 min data2. Do not use 14/16 min
Enable all SF, except the 3 Closed filters, (We know the CLoseD filters are good, and we are looking for alternatives) and roofinfilter1pole (redundant)
pf training back to 1.8 pf, pearsons 0.95
build systems the usual 50,000 / 250 way
export parm to excel.
sort on sf. Get most popular SF (alphabetical sort will do)
choose the most poplar sf from above (remove all other SF), put in GSB, and do the usual 50,000 / 250 systems
do stats on them

then choose the next most popular sf.... and repeat building 50,000/250 -> stats

You will find 1 (maybe more) much better than closeD

I suspect this SF is unique to CL and or energies, but im not sure.

RandyT - 30-1-2020 at 02:32 PM

Peter, do the following results reflect what you are seeing with indicator validation on CL and if so, how do you interpret the better performance of a single indicator vs 2 indicators? Is this a suggestion that we should lean toward system development with 1 indicator?

admin - 30-1-2020 at 04:06 PM

Peter, do the following results reflect what you are seeing with indicator validation on CL and if so, how do you interpret the better performance of a single indicator vs 2 indicators? Is this a suggestion that we should lean toward system development with 1 indicator?

Interesting.
We should try each others list to compare.
Here is my one indicator list.
One thing to watch is, the identical test. How much variation is there in each test.
One user who has spent months on CL, got his top 9 inidcators. My list has all his top 9, but more too.

Ive also found a second SF that is slightly better.
It was not in my top 250 systems.
We need to get stats of SF, like we have of primary filters.

admin - 30-1-2020 at 04:20 PM

HI Randy
did you have pf 1.3, pearsons 0.9. Your list is really different from mine

RandyT - 30-1-2020 at 04:26 PM

Peter, a number of questions:

1. What is the "IsParallel" setting you have under optimization?
2. What session times are you using for this run?
3. Can I assume that entry mode is "Compare1/2"?
4. Any changes to Indicator weights?

There are some somewhat significant differences in your results. Any chance you could share your optimization settings file?

RandyT - 30-1-2020 at 04:29 PM

Yes, settings were as you prescribed. Attached is settings file.

Attachment: Login to view the details

admin - 30-1-2020 at 10:34 PM

Hi randy.
Default weights. Compare 1 compare 2
Sessions times in my files I sent you.
ignore parallel. Feature for me only... More later from me

RandyT - 31-1-2020 at 01:31 PM

Secondary Filter analysis on single indicator run on 50k systems for CL data1:15min data2:30min

Peter, you did not mention in your prescription above how many indicators we should use for these SF analysis runs. I'll run this again on 2 indicators.

admin - 31-1-2020 at 04:05 PM

Secondary Filter analysis on single indicator run on 50k systems for CL data1:15min data2:30min

Peter, you did not mention in your prescription above how many indicators we should use for these SF analysis runs. I'll run this again on 2 indicators.

I used closeDmiuns and closedover

stats for secondary filters are nearly done too in next build of GSB

admin - 31-1-2020 at 04:32 PM

Hi Randy, this is really good what you have done.
I have the feature in GSB right now, but its buggy and doesnt work correctly.
I think we either use the top 9 (or a few more like in my post of my cl) or we do the entire list. A comparision of the two would be interesting

admin - 31-1-2020 at 04:35 PM

Randy, your ration is not correct (likely not critical) I should have a max of 1.

of the stats works is we use the fitness of the top 1/3 systems for top, and bottom 1/3 for bottom.

We ignore the systems in-between

RandyT - 31-1-2020 at 04:42 PM

Happy to share this Excel file if anyone else wants to use it. Updating data could result in needing to tweak this a bit. No promises. Formulas are designed to be a bit adaptive.

Peter, I did have some trouble mapping the filter names from what I see in GSB param results vs what I see in the configuration window. Comments welcome on the values that show "0".

Attachment: Login to view the details

RandyT - 31-1-2020 at 04:45 PM

Randy, your ration is not correct (likely not critical) I should have a max of 1.

of the stats works is we use the fitness of the top 1/3 systems for top, and bottom 1/3 for bottom.

We ignore the systems in-between

Ok, I was not clear on your formula.

admin - 31-1-2020 at 04:49 PM