admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
MISC TIPS
Im making a thread on useful tips.
See the c#/us number is 88. This means 88,000 calculations per 1 system that appears in the GUI. (131281 /1485)
If you get a number say 500++ its going to take forever to build a lot of systems.
It means you have the wrong market, time frames, settings etc.
The lower this figure, the better. But if you build on mult time frames, the figure will go up. If you tighten the metrics in performance filter, the
figure will go up.
The above test is on CL 29,30,31 min bars.
For example if you changed the secondary filter (SF) from closedminusClose
to closedminusCloseBPV (big point value) and the SFentrylevel input was not changed from say 0 to 100 to roughly 0
to 2000.
It means GSB on ES will be using 0 to 4 (4 = 100/50 {bpv of ES}) on the SF filter
The correct value for closedminusCloseBPV on ES YM, ER, NQ is roughly around $900, so 0 to 2000 is a wide but acceptable range.
Before long I want a separate input for SFentrylevelCloseBPV, as the above issue bites a number of users.
If you build systems, and fitness = 0 it means something is drastically wrong.
ie you have moc (marketonclose) time set before the open time of your chart
The fact that a fitness figure appears, doesnt mean it will give systems in the GUI. There are numerous performance metrics that have to be met for it
to appear on the GUI
The stus. Sum total 51 means 51 cloud workers are connected to this manager.
|
|
|
avatartrader
Junior Member
Posts: 60
Registered: 1-10-2018
Member Is Offline
Mood: No Mood
|
|
Thank you for pointing out the observation on the #C/US.
I just did a run on the ES using what I thought were pretty reasonable and appropriate settings, and yet I had a pretty high #C/US as you can see in
the attached screenshot. Is there anything obvious in the settings other than building on multiple series with more than one data stream that would
result in a higher than normal number of calculations?
Being an ES system, I was sure to change the Secondary Filter indicator to CloseLessPrevCloseD. Otherwise, everything was fairly standard. I was using
Nth mode and also Dates mode so I could exclude a certain amount of recent data from the build and validation.
For exits (not shown) I only used the MOC.
Is there some formula or something to try and approximate what the number of calculations per system will average based on the settings, or is it
something that cannot effectively be known until build time?
My speed was also really low for using 20 workers across 2 dual Xeon servers, but they are older X5500 series and I am looking to remedy that issue.
However, I am not sure that would impact the number of calculations per unique system.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
For es i prefer no secondary data streams. so try es 29,30,31
All else I see looks fine. You also might want to use all 2018 as out of sample via dates. A few months means little.
There is no forumula for c#/us. If you had a really slow pc, it wont affect c#/us. It will however affect how fast the systems will be generated
though.
I have dual xeon 29.xx for hire at $10 a day subject to availability.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Something I see on the cloud from time to time is people use all secondary filters (SF)
SF apart from closeD... do work , but my extensive testing on ES, CL etc shows nothing is better than CloseD...
There will be some rare exceptions. I think vix.x highest high wasnt too bad.
So if you choose all indicators, your systems overall will not be as good, and the c#/us ratio much higher (bad)
Indicator 29 roofing filter1 pole should always be disabled as its redundant compared to indicator30 roofingfilter.
This is not critical, but will make GA a little more efficient. It was left in for compatibility reasons.
So my default is to only use the CloseD.. SF. CloseDminus in SF is also redundant compared to closedBpv.
However if sf is not GA, but closeDminusCLose, watch the note at the top of this thread about SFentrylevel values.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
In the first post the number 2231471.35 was uses instead of 131281. Mistake now fixed
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Im seeing people build on 30 31 32 min bars, not 29,30,31
this is good to get right as say on es - there are 14,13,12 bars in a day. This stresses the oscilators but 13 is central
with 30 31 32 we have 13,12,12 bars in a day. Not as good.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Cant get any systems to appear.
cant get get any strategies
Check the fitness figure on right side. If its zero, you have something drastic wrong. IE market on close set to exit before open.
Go to a worker or the standalone, right click top/diag. The refresh. This will show the systems GSB is working on right now. It will take a minute to
bring in systems.
If you have say nth=1 set, you are going to get 1/2 the amount of trades. This may mean systems dont go through your performance metrics. You could
loosen these off.
Set slippage and commission to zero.
Check you have the correct secondary filter for the market.
Stock indices use a closeD filter, most other markets use
Check your data is correct. 30 min or 15 is likely to work best. You might have done 1 minute my mistake. ie 1 minute data with multiplier of 1
 
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Cant get any systems to appear.
cant get get any strategies
Check the fitness figure on right side. If its zero, you have something drastic wrong. IE market on close set to exit before open.
Go to a worker or the standalone, right click top/diag. The refresh. This will show the systems GSB is working on right now. It will take a minute to
bring in systems.
If you have say nth=1 set, you are going to get 1/2 the amount of trades. This may mean systems dont go through your performance metrics. You could
loosen these off.
Set slippage and commission to zero.
Check you have the correct secondary filter for the market.
Stock indices use a closeD filter, most other markets use
Check your data is correct. 30 min or 15 is likely to work best. You might have done 1 minute my mistake. ie 1 minute data with multiplier of 1
Your training or other metrics might be too tight. Reduce the numbers here.
Your data might be too short in years, or training period too short to produce say 100 trades
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Users often get into strife changing settings. ie make moc time before open time, so no trades.
This week a user go into strife, and I couldnt even see what they did wrong.
You can right click the left GUI, and reset to default GSB settings. GSB defaults need a tweak though. Like optimize price data is true. (should be
false)
App settings can also be reset. Tools app settings, right click - reset to default.
A bit more work will be done in this shorty.
The feature has been in GSB for some time. I just never knew about it.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
LET THE (GSB) FORCE GUIDE YOU.
We all have pre-conceived ideas and assumptions.
Its best to let GSB guide and challenge you.
For example, how many oscillators should you have in a system is a good example.
Opinions would be strong and varied.
Well I will show some objective ways to tell.
2 oscillators. GSB with 70 workers couldn't get over 1500 systems. (exclude duplicates)
The reason is say 40 osc * 40 osc = a max of 1600 possibilities. (This is very over simplified.)
So I built systems with 2 (1517 systems) 3 (30,000 systems ),4(30,000 systems ),5(30,000 systems) oscillators.
nth = no trade
nthday=80
ES 30 minute bars, verified on 25, 35 minute bars.
Training was set to 100%, pearsons 0.98, PF 1.8
Out of sample dates used 6.30.2015 to 2018.11.11
To speed the process up, I didnt re-verify the systems that passed verification on nth no trade - on nth = trade.
The point I can not stress more is we now have a powerful tool to guide as to what works best. This gives us objectivity and a benchmark for
any and every variation we make. If you just decide on different bar time, verification bars, entry operands - you are likely playing pin the tale on
the donkey.
You should validate your method.
This video from last year outlines it.
https://youtu.be/2R4t9uYzfD4
Note my default is now ES 30 minute bars, verified on 25, 35 minute bars, 3 oscillators. the UN-finished video in the private forum explains why.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Why not to use operators "+" with "*" or "/"
indicators are -100 to 100
so inidcator1 * ind2 can equal 10,000
if you add a third indicator ontop of this, it becomes 9900 to 10100
So whatever the value of the third indicator is, the results are not significant.
So you have build a architecture that now has redundant code.
what they do is
result=indi1*/+ indic2 */+ indic3
if Secondary filter=true and result >0 then buy
not only one operator of "*/+" is used at a time
So you can use +, or * or * * / but should not use + with * /
Future builds of GSB will lock this combination out.
|
|
|
admin
|
Thread Pruned
23-4-2019 at 05:11 PM |
![[*]](images/davis/default_icon.gif){kind=icon}
