| Pages:
1
..
15
16
17
18
19
..
47 |
Daniel UK1
Member
Posts: 470
Registered: 4-6-2019
Member Is Offline
|
|
Quote: Originally posted by admin  | Im having great success with secondary filters.
The top rating filters via secondary filters stats are NOT the best secondary filter.
Note my benchmark using the two closed SF vs some of the others on the list.
|
Hi Peter many thanks for sharing.
I am having a bit difficulty to understand what you mean with
"Top rating filter via secondary filters" ? and that those are not the best secondary filters?#
And
Note my benchmark using the two closed SF vs some of the others on the list."
I could not understand which one in your pic was the benchmark, and what is the "other" you arereferring to? and which "list do you mean?
Sorry for not understanding your properly.
Thanks
|
|
|
RandyT
Member
Posts: 123
Registered: 5-12-2019
Location: Colorado, USA
Member Is Offline
|
|
Peter, can you clarify for me what "CloseDminus a7d Closed" means on your graphic?
How are you identifying and choosing these other SF candidates since the parameter statistics sounds like it is leading us down the wrong path?
Can you also articulate how those values are calculated in the parameter statistics? Is this somehow using Fitness in the calculation? Or just
frequency that the indicators are used?
Thanks received (1):
+1 Carl at 2020-02-05 14:35:02
|
|
|
RandyT
Member
Posts: 123
Registered: 5-12-2019
Location: Colorado, USA
Member Is Offline
|
|
Thought I would share with the group a little work I am doing to further analyze indicator performance. The following graph is based from top 1000
fitness function values taken from a dataset of 50k systems. Clearly shows the indicator and sf-indicator combination that was most successful.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
| Quote: Originally posted by RandyT | Thought I would share with the group a little work I am doing to further analyze indicator performance. The following graph is based from top 1000
fitness function values taken from a dataset of 50k systems. Clearly shows the indicator and sf-indicator combination that was most successful.
|
That looks nice.
The critical thing is, even though highlowd was clearly the top, its bad when you build the 50k systems, and do that stats.
However a bit lower in the pecking order is some diamonds.
This is one factor to remember as we automate more of the entire process.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
On a roll this week. This was using RSI as SF in indicator selection, and RSI only as secondary fitlers.
this is the 91 top VSS
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
| Quote: Originally posted by Daniel UK1 | | Quote: Originally posted by admin | Im having great success with secondary filters.
The top rating filters via secondary filters stats are NOT the best secondary filter.
Note my benchmark using the two closed SF vs some of the others on the list.
|
Hi Peter many thanks for sharing.
I am having a bit difficulty to understand what you mean with
"Top rating filter via secondary filters" ? and that those are not the best secondary filters?#
And
Note my benchmark using the two closed SF vs some of the others on the list."
I could not understand which one in your pic was the benchmark, and what is the "other" you arereferring to? and which "list do you mean?
Sorry for not understanding your properly.
Thanks
|
So see here, a test of all SF (done with my favorite 19 CL indicators, # of indicators=3, training pf 1.8 pearsons 0.95, cl15min (data2=30 min)
It appears that CloseD are much better than all other SF.
 
but then we do the same test with no CloseDsf
We get this list
So we then build systems , but try HighLowD. Results terrible
Then we try close.preveHd
Results fantastic.
and we go down the list of all whats in green.
it gets better further down with some of the others.
Thanks received (3):
+1 moresi522 at 2020-03-10 09:35:17 +1 Carl at 2020-02-05 14:35:53 +1 Daniel UK1 at 2020-02-05 03:58:40
|
|
|
Daniel UK1
Member
Posts: 470
Registered: 4-6-2019
Member Is Offline
|
|
| Quote: Originally posted by admin | | Quote: Originally posted by Daniel UK1 | | Quote: Originally posted by admin | Im having great success with secondary filters.
The top rating filters via secondary filters stats are NOT the best secondary filter.
Note my benchmark using the two closed SF vs some of the others on the list.
|
Hi Peter many thanks for sharing.
I am having a bit difficulty to understand what you mean with
"Top rating filter via secondary filters" ? and that those are not the best secondary filters?#
And
Note my benchmark using the two closed SF vs some of the others on the list."
I could not understand which one in your pic was the benchmark, and what is the "other" you arereferring to? and which "list do you mean?
Sorry for not understanding your properly.
Thanks
|
So see here, a test of all SF (done with my favorite 19 CL indicators, # of indicators=3, training pf 1.8 pearsons 0.95, cl15min (data2=30 min)
It appears that CloseD are much better than all other SF.
but then we do the same test with no CloseDsf
We get this list
So we then build systems , but try HighLowD. Results terrible
Then we try close.preveHd
Results fantastic.
and we go down the list of all whats in green.
it gets better further down with some of the others.
|
Got you Peter, thanks.
This really shows that each SF indicator lets say in top 50% needs to be tested individually with 50k WF and get proper stats on, since this shows
that you can easily get fooled by the initial stats of SF test..
You propose that SF indicator tests needs to be done with SF Closed filters turned on and then off in general for a new market? this is among the
things what i take away from this at least.
Regards
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Hi Daniel,
We still need to revisit even ES again now we know this.
Also should try 15/30 minutes
It highlights the fact that we cant use the approach of "turn every option on at once method"
That was touched on in the last methodology video
All sort of things could be tried again on CL to, that we have tried in the past with SF closeD
I may do NG next
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Here is current CL system im working on. It used 5 separate indicators - which gives a little better result than 3.
However im not happy with 5 indicators in that the correlation of the top families with each other is very high.
  
Thanks received (1):
+1 LucaRicatti at 2020-02-12 19:11:27
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
updates on CL results.
Earlier total of 3955 (a few posts above) is now 9380.
2019.2.28 till 2019.12.31 results are g / h
Im still playing around with many options, but it looks outstanding.
Will publish more details when the dust has settled.
|
|
|
Daniel UK1
Member
Posts: 470
Registered: 4-6-2019
Member Is Offline
|
|
| Quote: Originally posted by admin | updates on CL results.
Earlier total of 3955 is not 9380.
2019.2.28 till 2019.12.31 results are g / h
Im still playing around with many options, but it looks outstanding.
Will publish more details when the dust has settled.
|
Looks very very good Peter.
What do you mean with
"Earlier total of 3955 is not 9380."
Regards
Daniel
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
| Quote: Originally posted by Daniel UK1 | | Quote: Originally posted by admin | updates on CL results.
Earlier total of 3955 is not 9380.
2019.2.28 till 2019.12.31 results are g / h
Im still playing around with many options, but it looks outstanding.
Will publish more details when the dust has settled.
|
Looks very very good Peter.
What do you mean with
"Earlier total of 3955 is not 9380."
Regards
typo, ive fixed it now thanks.
Daniel |
|
|
|
Daniel UK
Junior Member
Posts: 2
Registered: 19-1-2019
Member Is Offline
Mood: No Mood
|
|
Started to check in detail diff between real live trading and backtest, its done by hand.. usually compare and take note when i am restarting trade
server... number in white cell is number of trades pnl is calculated from.. Real is with IB commission and based on live trading and real fills,
Backtest is with IB comission only. Strats with gsb in name is as it says, GSB systems. Obviously amount of trades since i started tracking like this,
is not statistical sign, but perhaps interesting anyway..
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Im going to publish a bit more on CL systems tomorrow I hope.
Have got systems to trade live, and interestingly put a GSB system through cluster analysis with good results. I don't normally use cluster analysis
but thought it would be good to try.
By the way, the next build of GSB has 5 times faster file write WF for EWFO output.
Thanks received (1):
+1 cotila1 at 2020-02-21 12:59:14
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Im showing an optional validation methods at the end of the process - and later will show how we got CL systems
GSB exports into EWFO. EWFO is much more specialized in WF ability.
Fitness can be curve fitted like many other things in trading. So im going to take 3 years of data off and run fitness stability scores.
This takes some time as each fitness we try is a separate WF run.
The best fintess is area under drawdown. It has best metrics of astab-c,rstab-c, np & DD, etc
Incidently TS, fitness is half way down the rank (Np/-draw-down*winners)
I then put the full dates back to 20191231 and do the WF with fitness sum of area under draw-down curve)
Note since 2013 in sample is exactly the same as out of sample
a rolling wf also works well.
I find GSB systems work better with anchored WF, but feel rolling is a greater stress test.
Tomorrow we will show the cluster analysis results
Thanks received (6):
+1 uhrbi at 2020-04-05 05:17:01 +1 RandyT at 2020-02-23 16:31:27 +1 OUrocketman at 2020-02-21 16:05:47 +1 cotila1 at 2020-02-21 12:59:49 +1 Daniel UK1 at 2020-02-21 02:37:31 +1 Bruce at 2020-02-21 02:13:01
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
here is the results of cluster analsysis.
anchored
and rolling
It is very rare to find systems that have a perfect pass. Keep in mind there are additional criteria that EWFO has compared to TS WFO
|
|
|
gdk13
Junior Member
Posts: 6
Registered: 6-6-2019
Member Is Offline
|
|
Peter, your GSB is an awesome product ... and it only keeps getting better!
I have generated several ES, CL and S systems, which I will now run through TS to verify they are running properly. Then work with Portfolio Analyst
to choose non-correlated systems. Will add some position sizing and hope to be live trading in the next month.
Seeing GSB stats builds a lot of confidence -- especially when systems are able to hold together under stress testing on unseen data.
Thank you for all you've done and continue doing.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Big thank you gdk13 for your comments.
Here is another CL system.
It has (deliberately different) secondary filter to the above.
It passed all EWFO tests perfectly. (hard to do)
Interesting that the systems are UN-correlated. Thats not an easy thing to get when day trading the same market, time zone / bar interval.
|
|
|
moresi522
Junior Member
Posts: 50
Registered: 31-1-2020
Member Is Offline
|
|
Hi,
first af all I want to apologize for my english and for each grammar mistake.
I want to share with you my first complete experience of ES Methodology.
I have some doubts about results and I really would like to hear your opinion.
in the xls file attached you could find all tests I made with ES. I simply loaded the file that GSB give us with default installation, tuning
parameters as video "Genetic System Builder, updated methodology! Build a trading system with great out of sample result"
(file attached Baseline.gsboptset) and run for 50k systems. At the end I run Macro M1 and immediately I found inconsistencies with the video results
(see Baseline Sheet A50 -my fitness result- A52 - Peter's result). The fitness result are so much lower than video result. And I've only run run 50k
system as peter does with the same (I think) parameters.
As secondo step I looked for good indicators (file SearchIndicators.gsboptset). 20k Systems and macro IndicatorStats+FullSettingsReset_new.gsbmacro
under C:\GSB\Data\Settings\Macros gives my 9 indicators
Decycler,Oscillator,Dmi,DmiPlus,Momentum,RoofingFilter1Pole2,RSI,CounterTrend,CloseLessPrevHighD,VolumeLessPrevAverageFc
Thirth step. Then I'v used that indicators with all settings, the results are slightly better than Baseline (Sheet 3IndAll from A46), but nothing
special.
fourth step. If you look for 3IndCustom1 Sheet (file 3IndCustom1.gsboptset) you can see the parameter list I have used (a little different from video
indication, It's my mistake but I'd like to see how it worked). Fitness value for Main Astab e Vss are better but far from Peter's video.
At the end I've set all right parameters (well, I think), (File 3IndCustom1Final.gsboptset) but the result is poor and even worst from "not
reccomended" setting of 3IndCustom1 sheet. I'm definitely doing something wrong, but sincerly I don't understand where.
Another strange thing is that running macro 5 it found only 57 elements, not 90.
EDIT: it was a mistake, date field are set from 2000 to 2018
I have also all 250 WF systems if you want.
I am absolutely available to give you all the files and informations that I have.
What do you thing?
Thanks in advance for any advice
Attachment: Login to view the details
Attachment: Login to view the details
Attachment: Login to view the details
Attachment: Login to view the details
Attachment: Login to view the details
|
|
|
RandyT
Member
Posts: 123
Registered: 5-12-2019
Location: Colorado, USA
Member Is Offline
|
|
| Quote: Originally posted by moresi522 | Hi,
first af all I want to apologize for my english and for each grammar mistake.
I want to share with you my first complete experience of ES Methodology.
I have some doubts about results and I really would like to hear your opinion.
in the xls file attached you could find all tests I made with ES. I simply loaded the file that GSB give us with default installation, tuning
parameters as video "Genetic System Builder, updated methodology! Build a trading system with great out of sample result"
(file attached Baseline.gsboptset) and run for 50k systems. At the end I run Macro M1 and immediately I found inconsistencies with the video results
(see Baseline Sheet A50 -my fitness result- A52 - Peter's result). The fitness result are so much lower than video result. And I've only run run 50k
system as peter does with the same (I think) parameters.
As secondo step I looked for good indicators (file SearchIndicators.gsboptset). 20k Systems and macro IndicatorStats+FullSettingsReset_new.gsbmacro
under C:\GSB\Data\Settings\Macros gives my 9 indicators
Decycler,Oscillator,Dmi,DmiPlus,Momentum,RoofingFilter1Pole2,RSI,CounterTrend,CloseLessPrevHighD,VolumeLessPrevAverageFc
Thirth step. Then I'v used that indicators with all settings, the results are slightly better than Baseline (Sheet 3IndAll from A46), but nothing
special.
fourth step. If you look for 3IndCustom1 Sheet (file 3IndCustom1.gsboptset) you can see the parameter list I have used (a little different from video
indication, It's my mistake but I'd like to see how it worked). Fitness value for Main Astab e Vss are better but far from Peter's video.
At the end I've set all right parameters (well, I think), (File 3IndCustom1Final.gsboptset) but the result is poor and even worst from "not
reccomended" setting of 3IndCustom1 sheet. I'm definitely doing something wrong, but sincerly I don't understand where.
Another strange thing is that running macro 5 it found only 57 elements, not 90.
I have also all 250 WF systems if you want.
I am absolutely available to give you all the files and informations that I have.
What do you thing?
Thanks in advance for any advice
|
Moresi,
I am no expert, but here are a few comments about what I see in your Final settings.
1. You have an exit mode set. If you are trying to duplicate results Peter is showing, you will want to disable all exit modes.
2. FYI, in "Sign mode, I believe we are finding that SumOfIndicatorsSign is better than "Original".
3. In the Optimization section, if you are following Peter's new methodology, you will want to set Training 100%, Test 0%. I think this may be
limiting your training data given you are setting Nth Day = 1, Nth Day Period 80.
4. You'll also want to set Trading Periods - Post Build Auto Nth Day Mode = All.
5. In General Volume, make sure that if you have up/down volume tick data (2 columns for volume) that you set this to TickCount. It will matter for
some indicators.
6. Lastly, make sure you are running with latest indicators. These are available for download in one of the forum threads. There have been some fixes
to these lately.
That should get you a bit closer.
And BTW, no need to apologize for your English. I should be apologizing because I only speak one language... 
Edit: Link to latest macros: https://trademaid.info/forum/viewthread.php?tid=262&goto=search&pid=4649
Thanks received (1):
+1 moresi522 at 2020-03-08 13:13:46
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
| Quote: Originally posted by moresi522 | Hi,
first af all I want to apologize for my english and for each grammar mistake.
I want to share with you my first complete experience of ES Methodology.
I have some doubts about results and I really would like to hear your opinion.
in the xls file attached you could find all tests I made with ES. I simply loaded the file that GSB give us with default installation, tuning
parameters as video "Genetic System Builder, updated methodology! Build a trading system with great out of sample result"
(file attached Baseline.gsboptset) and run for 50k systems. At the end I run Macro M1 and immediately I found inconsistencies with the video results
(see Baseline Sheet A50 -my fitness result- A52 - Peter's result). The fitness result are so much lower than video result. And I've only run run 50k
system as peter does with the same (I think) parameters.
As secondo step I looked for good indicators (file SearchIndicators.gsboptset). 20k Systems and macro IndicatorStats+FullSettingsReset_new.gsbmacro
under C:\GSB\Data\Settings\Macros gives my 9 indicators
Decycler,Oscillator,Dmi,DmiPlus,Momentum,RoofingFilter1Pole2,RSI,CounterTrend,CloseLessPrevHighD,VolumeLessPrevAverageFc
Thirth step. Then I'v used that indicators with all settings, the results are slightly better than Baseline (Sheet 3IndAll from A46), but nothing
special.
fourth step. If you look for 3IndCustom1 Sheet (file 3IndCustom1.gsboptset) you can see the parameter list I have used (a little different from video
indication, It's my mistake but I'd like to see how it worked). Fitness value for Main Astab e Vss are better but far from Peter's video.
At the end I've set all right parameters (well, I think), (File 3IndCustom1Final.gsboptset) but the result is poor and even worst from "not
reccomended" setting of 3IndCustom1 sheet. I'm definitely doing something wrong, but sincerly I don't understand where.
Another strange thing is that running macro 5 it found only 57 elements, not 90.
I have also all 250 WF systems if you want.
I am absolutely available to give you all the files and informations that I have.
What do you thing?
Thanks in advance for any advice
|
I too got caught out running macro5,6 and getting less than 90/91 systems.
The reason likely is the dates are not set back to 2000-2015.
Run macro 11/12 instead of 5/6.
With the new indicators (beta mode)and some other tweaks, im getting better results than in the video.
In time the beta indicators will be added in non beta mode.
And I will update the docs.
basically to choose indicators on ES, follow whats in the CL docs
see step 4
https://trademaid.info/gsbhelp/Crudeoilmethodology.html
|
|
|
moresi522
Junior Member
Posts: 50
Registered: 31-1-2020
Member Is Offline
|
|
> The reason likely is the dates are not set back to 2000-2015.
you are right Peter. My bad.
I'm not able sto start Macro 11/12 Peter, I've downloaded the last macro from here
https://trademaid.info/forum/viewthread.php?tid=262&goto=sea...
but I got this error (see image attached) for both macros
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
that looks like m5/6 not 11/12
go to macro group2, and link the m5 to m11-wf_statsAstab
(you can unlink if needed by right click group2 - m5)
|
|
|
moresi522
Junior Member
Posts: 50
Registered: 31-1-2020
Member Is Offline
|
|
Sorry, It was M5. But I have the same issue with M11
EDIT:
Ok, It works. All macros MUST be copied under
C:\GSB\Data\Settings\Macros
Attachment: Login to view the details
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
for days like today where there is a limit down move,
adding this code to ES systems I feel has merit.
Im NOT adding it to GSBSYS1es but am to most of my other ES systems
the logic is, if the open has droped so much its a limit move (exchange wont allow more movement) there is decent chance market will go up, not down.
So dont go short.
I just implemented this minutes before the open.
For today it means dont trade shortif at 9am central time the market is <2910. (gsb systems wont trade long unless market goes miles up.)
I am asleep in a minute and will not reply till after the close on this issue
 
|
|
|
| Pages:
1
..
15
16
17
18
19
..
47 |
![[*]](images/davis/default_icon.gif){kind=icon}
