| Pages:
1
..
43
44
45
46
47 |
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Just some ideas for experienced users.
Happy to say I learn things from the user base.
One idea I think is valid, is to build systems on one day of the week only (or more). I saw a system doing this, that has 600 or so trades, and good
out of sample. Results on the other days terrible.
The other idea which is more complex is to build long only systems, and put on the same chart as short only systems.
The weakness is you get 1/2 the amount of trades when you build your long only system, so the sample size is smaller and that means its likely less
robust.
One smart user as combining a long short system, that is much better going long, and disabling short entries
I like this idea better then long only & short as its got more trades and might be more robust as it can trade long and short
However its very valid to make long only systems, esp on stock indices
Then he gets a long short system that is stronger short, and disables long entries, and combines them on the same chart.
IM not going to give more detail on this, as its complex. GSB has long short ration in the GUI
However the day of week ideas Im going to test and build the features into GSB automation
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
A number of significant updates.
1) Egg on my face. A lot of my later testing on gold accidentally had the dates changed to the out of sample dates, not the in sample dates.
That means much of my work in the last month is not correct. Im rapidly redoing things and it was not all wasted.
Lession learnt, make copies of a master opt settings files and always make changes to this file. Thats to stop the human error that I made
2) I have a short cut method that reduces the amount of cpu but a bit over 50% with no significant negatives and its really simple to implement.
I will document this in the next week.
It is valid to make trading systems on a day of the week, as long as you have >300 trades (rough number, more the better)
GSB and GSB automation and automation excel spread sheet can now do this.
It may be some time before I try to build systems using this due to time pressures
However here as an example. (picture coming later) You need todays GSB 66.25, automation 20230814.3 and excel build version 67)
Automation is shuffling a day of the week each test.
One comment, the pearsons and number of trades in test period needs to be reduced significantly as you will always get lower pearsons, trade count and
np/dd when you have in effect 1/5 of the data (one day a week)
Later im going to try trading 4 days a week and excluding 1 day
Repeatability tests
You can now put 100's of tests, and see the results of each 100 tests all in one spread sheet. (takes excel 30 min or so to process)
Example shown here was lots of 10, instead of 100
Thanks received (3):
+1 Bruce at 2023-08-18 15:22:03 +1 Piet at 2023-08-16 00:57:54 +1 Carl at 2023-08-16 00:49:34
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
An update on repeatability testing. Important as some of what I wrote in the last few weeks was wrong due to error in config file.
Here is a summary
the short cut method has no negatives but saves 50% of cpu time
Repeatability when you do 100 tests (of 2500 each) is about 4.5%
this means any change from your benchmark that is not improving by greater than this amount should be disregarded
The green10Fav300 macro was not a good idea. It gave a little worse results
ON the current GC setup,
3 indicators, 4 indicators with multiply or divide all gave the same out of sample results. Therefore 3 indicators with multiply will be by default
I will publish spread sheet of results in a few days
Thanks received (2):
+1 Bruce at 2023-08-24 00:26:36 +1 TwntySQ at 2023-08-23 02:35:40
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
here is the output of todays build of gsb automation spreads sheet
the repeatability figure of 4% is cell ac12 on the sort tab.
In gsb automation i will add version 67.2
lots of work on repeatability and massive improvement in calculation speed
Attachment: Login to view the details
|
|
|
TwntySQ
Junior Member
Posts: 26
Registered: 17-2-2023
Member Is Offline
|
|
Hi,
Some thoughts from a norwegian with an increasing interest of algorithmic trading...
I recently came across a video talking about genetic algorithms and one of the fitness functions that was mentioned there was the Martin ratio, or
more commonly known as the Ulcer index.
I don't have GSB in front of me now, but I cannot remember seeing the Ulcer index as a FF.
In short terms as of my understanding the Ulcer index measures the depth and duration of a drawdown, and rewards shallower and shorter/lesser
durations of drawdown.
Paper describing Ulcer index : https://www.tangotools.com/ui/ui.htm
Youtube video that describes this with an genetic algorithm. The results are not the point here, but the content of the Ulcer index. I find his
thoughts on elitism, parent selection and generation to be intriguing.
https://www.youtube.com/watch?v=2XQ3PsZActM&ab_channel=neuro...
Note: My understanding of genetic algorithms and algoritmic trading is nothing to brag about, nor my knowledge of computer science 
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
@TwntySQ
good to hear from you & this is worth discussing.
We already have ave distance between new highs, ave of 5 worst draw downs, pearson correlation to a straight line (one I use a lot, but never in
system building)
Despite these and many more, I now never use for system building anything but np*at and recommend others to do the same.
IN GSB automation you can try any fitness you want, to see the results. Keep in mind after the repeatability testing that ive spent two months on, you
need to do at least 100 * 2500 tests
If you fail to do repeatability testing, and your looking for small improvements, you end up with false conclusions. As this is foundational, it was
well worth the massive effort I put into this.
If you greatly want ulcer, i can add it but I think its not going to be helpful
Thanks received (1):
+1 TwntySQ at 2023-09-20 02:04:00
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Here is high level of a month or so of research done on Gold and Nasdaq futures
using the 500 * 2500 tests short cut method of 2 build sof systems per one indicator build
using 82 trial mode indicators
nth mode 10 clearly is best on both NQ and GC as far as repeatability of results
ON NQ out of sample results were consistent regardless of nth value
on GC results were more erratic according to nth value with no clear pattern, however nth 10 was among the best
repeatability of one lot of 100 tests were all within 3% of each other with 87 indicators, nth 10.
With 192 indicators (more that whats released even to beta users) repeatability was 4.67% for gold and 8.42 for nq
thanks greatly too all those who contributed CPU power to this project. Likely the most CPU intensive project I have done. Really tedious to do the
research but exciting with the end result.
MORE DETAILS HERE
https://trademaid.info/gsbhelp/ImprovedRepeatabilityMethod.h...
Thanks received (1):
+1 TwntySQ at 2023-09-22 03:53:16
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Here is an example, that in once sense could be any market, but is NQ using 23 hours bars, 5.5 days a week, long only
With massive effort i got 3 extremely linear systems that are very different to all other GSB systems built. It has a little non gsb code that
improved performance too.
Effort is also high as there is about 4 times more bars than a standard 30 min 830 to 1500 system, so cpu time is longer and gsb workers are using 10
to 15 gb of ram each
I will publish more as time permits. This also makes optimization in TS significantly slower. (I nearly always do manual optimization checks in TS
with EWFO if I sell a system)
there are only 2500 systems per test, but two tests per indicator build (short cut method I wrote about earlier)
GSB needs to new features to maximize this sort of system
One of them is break even stop.
Im also working on getting atr stop and a new filter.
Thanks received (8):
+1 Ketil at 2023-10-23 02:19:57 +1 TradingPrice at 2023-10-21 13:07:29 +1 JozefSusko at 2023-10-21 09:22:37 +1 autotrader2023 at 2023-10-21 08:36:51 +1 thowoc213 at 2023-10-21 02:58:12 +1 Piet at 2023-10-21 01:46:50 +1 Bruce at 2023-10-20 21:51:53 +1 SwedenTrader at 2023-10-20 19:18:42
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Feedback, were people more happy about the gsb features, or the 23 x 5.5 systems?
thank this post for gsb features.
next post for 24 x 5.5 systems (fine to do both)
atr stop has been coded, but i dont have the release yet.
Thanks received (2):
+1 Bruce at 2023-10-24 15:16:35 +1 TradingPrice at 2023-10-23 20:20:08
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Feedback, were people more happy about the gsb features, or the 23 x 5.5 systems?
thank this post for 24 x 5.5 systems (fine to do both)
Likely 2 to 4 weeks work to get ideal setup for 23 hour 5.5 day systems. Very cpu intensive.
Im doing NQ now, and then need to do ES.
Will be interesting to see if the setup is identical.
Thanks received (4):
+1 jasonp111 at 2023-11-11 11:55:51 +1 SwedenTrader at 2023-11-01 11:20:56 +1 Bruce at 2023-10-24 15:17:07 +1 TradingPrice at 2023-10-23 20:20:22
|
|
|
portfolioquanttrader2020
Member
Posts: 269
Registered: 30-4-2020
Member Is Offline
|
|
ATR
Quote: Originally posted by admin  | @daniel
here is gsbsys2nq
Only a filter is being optimized, no other settings are touched, so perhaps not a perfect example
Basically there is a 18% reduction in the amount of trades and a $500 in total increase in NP
equity curves are very parallel
Note that in run 2 and 4 only, the input value of 0.2 made this filter pass all trades
but when the filter kicked in in most other runs 0.06 was a fairly consistent value
atr value of 0.42 and above pass 100% of all trades
|
HI Peter where is atr value?
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Here are all the files needed to build nq systems on 23 hour days, 30 min bars + sunday session
Note workers will be ram hungry. Approx 10 gb per worker and build time will be slow as there are roughly 3 times more bars than 830 to 1500 30 minute
session
I will document this as time permits
These systems typically are close to new highs
Changes I do on final system once its in TS.
A break even stop should be added of $500 to $1800 range. (not in GSB yet)
Profit target increased up to $6000
Attachment: Login to view the details
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
updated file
UNZIP this file, and then click on Click-on-this-to-install.bat
Attachment: Login to view the details
documentation https://trademaid.info/gsbhelp/BuildingNasdaq23hrX55daySwing...
Thanks received (4):
+1 Bruce at 2023-11-05 20:29:36 +1 TradingPrice at 2023-11-02 19:23:04 +1 Carl at 2023-11-02 10:17:07 +1 coolj010 at 2023-11-02 04:25:21
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
[/rquote]
HI Peter where is atr value?[/rquote]
I cant find this code sorry.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
The long only 23 x 5.5 nq systems ideal environment is a over sold market that has long term reversals
On a monthly basis, all 3 of these are at new monthly highs. (1 at new daily highs)
More research has show
20 &60 min bars work just as well.
20 min will be slower to build, need even more ram but give better liquidity as 2 of 3 orders are not at the end of 30 min bar.
60 min is faster, but means all orders are at the end of the 50 min bar, which is not good for liquidity
Whats also different compared to some other GSB systems is the DailyClose price which is used is at midnight or 5pm sunday while GSB intra-day systems
that use closeD (not all of them do)
use a 3pm close.
Couple that with the fact that they are longer term systems, means you have got some good diversification.
enclosed are screen shots of my first systems. These are basically the same system but 2 nq, 1 es, 1 nq is 23 hour day, 1 nq is 22 hour day and
parameters differ a little.
  
Thanks received (5):
+1 cotila1 at 2023-11-24 10:18:18 +1 Daniel UK1 at 2023-11-15 05:00:20 +1 TwntySQ at 2023-11-14 12:45:07 +1 TradingPrice at 2023-11-11 03:40:16 +1 thowoc213 at 2023-11-09 03:16:36
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Here is a system a GSB user sent me this week.
Here where it differs from what ive spend much of the last month on
Its '@es.d 830 to 1515 30 min data while i have been on @nq 23 hour 5.5 days 30 min data
I think all else is the same. Profit target, stop, breakevenprofit lock, secondary filter
Note also how the last few years are the best in history since 1997
out of sample since February 2022 and all years before 2007
 
|
|
|
TradingPrice
Junior Member
Posts: 21
Registered: 4-8-2022
Member Is Offline
|
|
@admin
regarding default settings for the new 23 x 5.5 methodology, it uses 30 min NQ data with the session called "MOC @ 23:59"
well that didn’t make sense to me as GSB doesn’t see 23:59 with 30 min data provided
so I've re-backtested same systems with 30 min bars from 1 min data, and surely got different results
but when changing session end to 23:30 you get same results in GSB
** best match I got from GSB to tradestation is when specifying 23:30 as end of session for same systems made with default settings ( 30 min bar data
+ session MOC @ 23:59)
so basically, testing with 30 min data + MOC @ 23:59 = testing for daily close price @ 23:30
I might be wrong, but I thought I'd post as subsequent tests that were done with 1 min data aren’t exactly the same (20 & 60 mins bars)
overall, fantastic work Peter, it could very well work on other markets as well
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Im getting exact matches. use 24 x 7 session time not 23:59. (my settings I gave people were @23.59 but i got it to match. apart from the time field)
make sure your ts chart data and gsb data start at the same date.
if your stuck, we can run the match check diags via anydesk.com
|
|
|
REMO755
Member
Posts: 181
Registered: 11-4-2021
Member Is Offline
|
|
Hello Pedro
Where are the macros to make the complete methodology?
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
@remo
they should be in the macros folder of gsb
C:\GSB\Data\Settings\Macros
This is the current macro im using.
Nothing has changed of late, except that im not looking at smaller periods oos of period b c d e, just in sample period A and all OOS period F.
why im not doing bcde is it takes a lot of cpu time and i dont really care much on a specific period, just the entire out of sample period
Attachment: Login to view the details
Attachment: Login to view the details
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
a few minor things
the session i used for building 23 x 5.5 day systems was not correct
i used one that end with 23:59
It should match whats used in ts chart.
It wont affect results massively but they wont match gsb to ts for this reason.
There is a bug in the hull moving average filter. While it improves the result using it, i suspect the end of session was closed, not the midnight
close.
Bug only shows on 23 hour data, not day session data. I expect the fixed version to have further improved results
Much of the week was spent on 23 hour nq systems, and I learnt a fair bit and made numerous variants of the same system that give significantly
different results
see correlation table
 
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Coming soon
I have made significant break through on crude oil day trading systems, though much of the concept is from another GSB user.
CL systems are fast to make, and seem to have high chance of having good out of sample results in unseen data
Here is the original system
the wf is another system that I built today.
Likely in the next release, I will include the default settings for these cl systems.
You must be on the 2024 may GSB resource manager to get this update. (in private forum)
 
Thanks received (4):
+1 TradingPrice at 2024-05-18 14:16:58 +1 coolj010 at 2024-05-17 16:55:24 +1 Carl at 2024-05-17 08:42:26 +1 Ketil at 2024-05-17 03:21:49
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
An update on what we are working on. Crude oil day of week systems.
This is pre release but you can see whats in the pipeline
https://trademaid.info/gsbhelp/Crudeoil2024dayofweekmethod.h...
Thanks received (1):
+1 TradingPrice at 2024-05-21 15:23:29
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
I have learnt a lot this week when it comes to weekly patterns. IE day of week.
I still have a few days more processing time to complete my research, and what Ive learnt is going to save others a lot of wasted time.
GSBautomation next release now has all the DOW features.
ie you can do mon,tue,wed..fr long only, then short only, then long and short
as well as optional all the above tests with one DOW forced to be True.
so what I did was WED forces true, then added all sorts of combinations of other days as I already new CL Wednesday Long&short was a good and valid
combination
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
SO much has been happening.. hard to get time to write it all down.
GSB users will range from people who run default settings, to those who via GSB automation run exhaustive checks on new markets, or re-explore old
markets.
There are many secondary filters that can be tested, and an exhasutuve test via gsb can be done, but its very cpu intensive to do.
I have a hybrid appoach i used on CL.
I built CL systems with a very good setup. Recorded the top 10 indicators of what was a good run.
Then I built 10,000 systems with these same 10 inidcators, but all secondary filters enabled.
Here are the results
From this I take the top ones, and test one at a time in gsb automation
I think CloseLessPrevCloseDBpv is the best SF for the day of week CL systems (a significant improvement I found this week)
but my previous best sf CloseLessPrevLowDPrevCloseD is number 76 on the list
Doing this is best left for the GSB user who knows GSB well, and has lots of CPU power. Eventually what is learnt will make its way into the default
settings
GSB automation has been updated today to allow import of SF from GSB into GSBautomation sf list
Thanks received (1):
+1 NickW at 2024-05-30 13:31:04
|
|
|
| Pages:
1
..
43
44
45
46
47 |
![[*]](images/davis/default_icon.gif){kind=icon}
