GSB BETA BUILDS

orginal /optimization fitness %
this tests all wf fitness compared to the orginal fitness.

This is experimental and i short term dont have time to test how valid or useful it is
use random space 1000 iterations wf muti threaded +- 25% range

Let me know if any users want a copy

next build
wf optimization profile

here is the same system with search space full (the max of all ranges allowed for each parameter in gsb. Note that the end results is both much lower and much higher. This implies that the final parameters of +-25% might still not be the optimal ones. Beware however that looking for the peak fitness as your only criteria for parameters is not a good idea.

I hope to have a later build that exports results to a csv file.
It would be good to map the top 1/3 of your say 300 systems, look at the profiles,
then compare to the botton 1/3 of your 300 systems.
Then change dates to be the years out of sample not seen, and see if there is any correlation to the curves. Ultimately we need to automate this... but before we do that we need a critera of what we want to see in a curve. This must be decided by the GSB user base




Attachment: Login to view the details

67.15x released
Has new rego where you the end user can label a machine. This is useful if you ever need to revoke a license and move it to a new machine, you now will know what one to revoke

Quote: Originally posted by admin

67.15x released Has new rego where you the end user can label a machine. This is useful if you ever need to revoke a license and move it to a new machine, you now will know what one to revoke

@TP
The files should work on builds of GSB, but the update has been problematic because if a manger or worker is open when RM updates, the files are locked and you get miss-match in versions, which will cause failure to miss match fixed.

RM does that already, but much more often than once a day.
as soon as a job has ended, they are restarted.
On the subject of RM.
Is this of interest to anyone?
It will only work on supermicro servers that are not VM

Whats hopefully coming in GSB - not finished yet
Option to force data1, data2 etc on specific indicators or filters

67.21 released to beta testers
has option to use datax on any indicator IE data2 and force it to only use data2
this feature is not at all tested, and may have minor compatibility issues with future build if we allow it say to have data2 or data3

the first build with stop entry and limit entry is getting close. Currently its buggy.
This will be a work in progress as there are many types of stop entries to be added.
ie buy at closed(1) stop, but at opend(0) stop + atr, but at pivotpoints stop etc

Last week, collectively we hit the highest programing hours in GSB that we have had for years. However much of this is back end data base work.
Here is whats coming soon
Adjustable parameters for filters
improved database handling of walk forward jobs.


Whats hopefully a little after this is, major re-write of how data base works.
What this will mean is improved worker performance, but new managers will NOT work with old workers
SO if you have a unfinished job in GSB automation running, set this to false closer to the time.
I will give plenty of warning about this.

A lot of work continues to go into GSB back end data base.
Big re-write of logic, and much of the logic not in the data base is self, to give big speed improvements.

An issue thats been with GSB from day 1, is if you have 30 workers running, only 10 might run. This doesn't matter greatly as if your CPU is not maxed out, GSB resource manager will just open more.
However this issue seems dramatically improved, perhaps 100% fixed in 67.72 onwards.
What this means is your limit of how many workers in GSB resource manager should be reduced, as well as amount of workers you start on the group startup.
on a 72 core machine, im getting cpu maxed out with 8 workers. Im not sure the number in previous builds, and it could be related to the settings im running, but it used to be higher than this.
This however is good and in perfect world, one worker would max out your entire machine. (this will never happen with GSB due to ram cache being used)
However another factor is that windows server 2022 works better than server 2019. The context of this is a machine with dual CPU's. Its an operating system limitation, but most apps cant span >1 cpu. I suspect server 2022 swaps what physical CPU an app is on, if the other cpu is not as busy.
If you have one physical CPU, i think the OS version will not make any difference. (small chance it will for users with > 32 cpu cores)
(the exception is windows11, prioritizes front apps much more than back ground apps, so is much slower at running workers unless GSB resource manager is in tiled mode)

68.17 released today
next build
macro options can load systems from a folder. Why I did this is I want to benchmark various cpu's to see the speed of nth macro and dates changes.
FixedRemoveDuplicatedIndicators for trial and purchased users.
A new build of GSB automation has been released. It has extensive day of week programmability.
Good week trading. one of the gsbswing60 hit two $520 / $5200 profit targets in < 12 hours. Thats a first ever from memory

next build of GSB
monte carlo wf.
Example here only 100 tests done (not a lot) but you can see many wf gave better improvements and many poorer - which is fine.
what would not be fine is if most would give much worse results
This test was also good as paramater range was +- 100% which is very wide.

also 2 new indicators for me to test and release if I feel they work well

There is more work to be done on refining the validation groups before I release the next build of GSB

This situation reminds me of when i was at a camp, and was showing the audience the flying tea bag trick. (light a empty tea bag with a match)
People were in hysterics, not laughing with me - but at me. It failed terribly and turns out only some brands of tea bags fly.
Well the montecarlo SR ration has a significant effect to predict out of sample results, as does verification.
However lowSR gives the best out of sample results, not high sr. No idea why and this may take some time to work out

what was done here
build 2500 systems using 1 pass indicator (not 2 pass method) - as i wanted a significant amount of systems that failed. However the setup nq short only was so good, even the bad systems were pretty good

get the top 300 systems (by macro) then do family mode to make only unique systems. We now have 233 systems

verify systems with 5,10,15,20,25,35,40,450,55,60 minute data and look only at the vss score of all the verification totaled together. result (ie -20% degradation is good, result like -70% is bad) (ignore exact figures)

then do mc wf (singled thread fastest with cpu cores for wf set to the amount of cpu cores we have)
take lowest 70 systems mc sr ratio in fav A, top 70 systems mc sr ratio in fav B
Do stats a, stats b

then take lowest 30 systems vss score (this is bad) of the worst 70 mc score, and compare to the highest vss score (this is good) of the top 35 of the 70 systems


bottom line, big difference in results between the two tests, and greater difference when we apply both filters


build 68.56 or later is needed to do this. I will release it soon

im feeling better about mc sr.
note the worse mc sr and best mc sr
low mc sr has move of the curves above zero

mc 1000 interations and verification gave a whopping 44% increase in out of sample fitness of the worse vss and worse mc sr vs best vss and best mc sror
or 36 % of the bottom mc sr figure (vss ignored) compared to the top 70 mc sr and the top 50% vss of those systems
36% derived from (27164-20179)/20179

68.64 pushed to beta testers
Significant change to lots of things in Walk forward. Monte carlo wf metrics (but not all the lines) are now done on all WF, and can be done via cloud workers
Numerous MC wf metrics, that I still need to test which ones work best.

Wishing you all a great Christmas and new year.
Its been a good year in trading profits and GSB development - with the compounding effect of positive returns starting to feel like its kicking in more.

Quote: Originally posted by admin

68.79 released to all users. There is a lot in this build with new settings too. Feedback welcome

I can't run version 68.79, I am getting the following message:

Quote: Originally posted by Ketil

I can't run version 68.79, I am getting the following message:

next release of GSB, macro to save all systems in .csv file.
Recommend you use the family command first to remove duplicates
File will be in C:\GSB\Data\Exports\Systems

i have all but 3 of the new normalzation modes working. Initial testing showed its going to be hard to beat centerscaling2 and highlowest2. But I will know more in 24 hours or so. The new modes have been added into GSB automation next release.

Also working on buy/ sell entry stop. I expect this to be good on NQ and gold markets. Initial testing showed it worked well on NQ, but GSB TS code results are miles out - so this has to be fixed to confirm

Ive had very good success with the default NQ day trading settings, with new features of entering on stop of highest (high,x)
I also felt that the current default normalization mode Centerscaling2 (followed by second the best method highesthigh2 was the best.
However when I look at further out of sample results 20141112 to 20150515 other methods have done very well. Results include $5000 profit target as I dont want the $30,000+ trades of 10th of april to allow GSB to curve fit to this day.
Enclosed is an example of one system I picked, that had superb out of sample results.
Many but not all systems did well on the Trump tariff era of feb march 2025 which hammered many - but not all systems. If you look in excel, period G result is np/dd of 2024111 to 20150415. 28 trades were typical, so np/dd of 2 over this period is acceptable, esp when you consider any 30,000 wins are removed by the 5k pt.
In excel what we care about is the f medium result and the amount of systems in Favorites D. (all of this done by GSB automation)
I will push this build in about a weeks time. If anyone wants to pre-test it, it may be ready this Tuesday

Here are some nq result using various entry types. More are being added. Whats clear is buy at highest(high,x) is much better than buy at market.
what we care about is F results, systems in Fav BCD, and G result (next out of sample period)

A very successful week building systems, and continuing to add more pivot entry types.
I worked methodically on the original gold settings using GSB automation, and got significant improvement. (The old settings were not great)
Then, an accident. changed one setting I should not have, - and no idea how I did that. Pure human error. Very significant improvement....

then a long time user who got my second gold system, sent it back but changing data1 to silver. Zero changes in settings, but zero slippage used here.

also, a trial user who I gave GC settings to came up with this. Perfect wf were oos, in sample and original system results all identical

Next build being tested for bugs, and still waiting on my entry types

You can also see on my first 2 gold systems, low correlation looking at the charts

There are now 57 pivot type of entries if you include the stop / vs limit as a different type.

The Increase in out of sample median has increased on gold systems by 42% compared to earlier build of GSB.
The amount of systems in Favorites D has over doubled!

One of the newer entry types (not shown) is even better. Still doing more tests.

Other markets also confirmed the same thing.
Very exciting developments.
Very early days yet, but Gold systems out of sample is good. I will update you when im ready to release the build. Not to far away I hope.

Quote: Originally posted by admin

Another significant week of development Next release (not this week) CL day trade settings update. Just as with nat gas and gold, new GSB has very significant increase in out of sample results. The devil is in the details I had a long term user amazed at the improvement of one setting. The secondary filter for stocks / indices has to be non ga/non normalized) closelessclosedBPV and for energies it has to be normalized (ga) closelessclosedBPV what i was not aware of is the filter (if used) closelessclosedBPV is NOT normalized. Next build we will have the normalized option as well. This means you can combine a closelessClosedBPC and other secondary filters of your choice. We have had for many years the option of data1, data2 etc. Here is the problem. We could build using data2, and wrongly think results have improved. Problem is results always vary significantly from one run to another, thats not bad and is documented elsewhere. However your systems may never have even used data2. This can actually be seen in gsb automation spread sheet stats. However we now have the option to only allow systems that use multiple data streams. Any that dont are never even tested. That way we now are sure we can see the effect of having at least one indicator have data2.