kiwibird
Junior Member
Posts: 10
Registered: 25-7-2017
Member Is Offline
Mood: No Mood
|
|
Methodology of GSB
I probably missed it but what methodology or algorithm does GSB. Is machine learning used on last n bars of data(and indicator values)?
I include a PDF of another method that you and the forum might be interested in. Your comments are welcome.
http://www.itrac.com/paper/itrac.pdf
|
|
|
admin
Super Administrator
Posts: 5069
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
It is machine learning, by genetic algorithm. GSB looks for consistently profitable algorithms with optional training, test, validation, optional
multi time frame and optional multi market verification, and strongly recommend walk forward testing.
The article looks like a good read, but will take time to go through. Im on holiday till late this week so working less hours.
|
|
|
JasonT
Junior Member
Posts: 61
Registered: 6-6-2018
Member Is Offline
Mood: No Mood
|
|
Training, Test, Validation
Hi Peter,
apologies for what might seem a simple question to you, but I'd be grateful for a bit of help. I understand walk forward optimisation, in sample and
out of sample optimisations. I'm trying to get my head around that in context with the training and test sections of the equity curves. So let me
see if I've got this right:
-GSB combines 3 to 5 different indicators out of a pre-loaded set of 36 or so, plus any extra customs ones a user loads up or you may include in the
future. Lets say a set of three like: MACD, RSI and CCI for example.
-It then adjusts the values of these indicators (eg Fast or Slow period) within that set over their various permutations and combinations (or a sample
of them) and for any systems that meet any filter criteria, these are displayed as unique systems in the table in the lower centre of the screen.
-After it has finished with that set, it replaces say the CCI with another indicator, like a stochastic, and then off it goes again.
Please correct me if I'm wrong. 
What then is the difference between the training data, the test data the validation data and the walk forward optimisation that is able to be
performed on those unique systems.
Thanks very much.
I'm enjoying the videos, working through the documentation and trying different aspects of the GSB to see what they do.
Best regards, Jason.
|
|
|
admin
Super Administrator
Posts: 5069
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
Quote: Originally posted by JasonT  | Hi Peter,
apologies for what might seem a simple question to you, but I'd be grateful for a bit of help. I understand walk forward optimisation, in sample and
out of sample optimisations. I'm trying to get my head around that in context with the training and test sections of the equity curves. So let me
see if I've got this right:
-GSB combines 3 to 5 different indicators out of a pre-loaded set of 36 or so, plus any extra customs ones a user loads up or you may include in the
future. Lets say a set of three like: MACD, RSI and CCI for example.
-It then adjusts the values of these indicators (eg Fast or Slow period) within that set over their various permutations and combinations (or a sample
of them) and for any systems that meet any filter criteria, these are displayed as unique systems in the table in the lower centre of the screen.
-After it has finished with that set, it replaces say the CCI with another indicator, like a stochastic, and then off it goes again.
Please correct me if I'm wrong.
What then is the difference between the training data, the test data the validation data and the walk forward optimisation that is able to be
performed on those unique systems.
Thanks very much.
I'm enjoying the videos, working through the documentation and trying different aspects of the GSB to see what they do.
Best regards, Jason.
|
Systems are built on training data, then you optionally look at test and validation data.
GA will mutate the systems and try different combinations to improve fitness.
You may have mixed up WF with genetic build of systems.
You were correct till you got to after its finished...
WF will get the best parameters for a section of data, then apply those parameters to the next 'unseen' bit of data. This loops till the 10 WF runs
are done. Oscillator values will change, but not oscillator types.
There is a lot to get your head around, so its no surprise you dont understand it all. Thanks for your comments on the videos. I have employed someone
to update & improve the documentation, but this will take a bit of time to do.
|
|
|
JasonT
Junior Member
Posts: 61
Registered: 6-6-2018
Member Is Offline
Mood: No Mood
|
|
Some questions about IS/OOS 'testing' sections of GSB.
Hi Peter,
I understand that Training% is the amount of data that is used to optimise/maximise GSB fitness. That data is In Sample (IS). Is
this correct?
Test% applies the optimal parameters found in the prior Training% IS data to unseen Out of Sample (OOS) data. There
is no further optimisation/maximisation/changes in the parameters. So if the equity curve continues on a similar trajectory with similar performance
attributes, then this shows the strategy has some resiliency and is expected to continue to perform similarly in the future (although it may not). Is
this correct?
If the Validation% data is also OOS, what is its purpose? What additional value does the method expect it to add?
Note that I've got some additional questions about Walk Forward optimisation that I'll ask next, but I am assuming that building strategies using this
Training/Test/Validation approach to IS and OOS data is a step prior to conducting walk forward optimisation on particular strategies that you like
the look of (as shown in some of your videos).
Many thanks,
Jason.
P.S. I'm very glad I've purchased the software and looking forward to mastering this incredible capability you've developed. Thank you.
|
|
|
admin
Super Administrator
Posts: 5069
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
| Quote: Originally posted by JasonT | Hi Peter,
I understand that Training% is the amount of data that is used to optimise/maximise GSB fitness. That data is In Sample (IS). Is
this correct?
Test% applies the optimal parameters found in the prior Training% IS data to unseen Out of Sample (OOS) data. There
is no further optimisation/maximisation/changes in the parameters. So if the equity curve continues on a similar trajectory with similar performance
attributes, then this shows the strategy has some resiliency and is expected to continue to perform similarly in the future (although it may not). Is
this correct?
If the Validation% data is also OOS, what is its purpose? What additional value does the method expect it to add?
Note that I've got some additional questions about Walk Forward optimisation that I'll ask next, but I am assuming that building strategies using this
Training/Test/Validation approach to IS and OOS data is a step prior to conducting walk forward optimisation on particular strategies that you like
the look of (as shown in some of your videos).
Many thanks,
Jason.
P.S. I'm very glad I've purchased the software and looking forward to mastering this incredible capability you've developed. Thank you.
|
Hi happy you are happy and wish you the best with GSB. Your are on of two users I have meet in person.
Validation gives you the human, the change to look at another out of sample period. This is fraught with danger as we the human will abuse this and
just go for the best results over the entire period. There would be merit in making gsb not sort or include validation in its default metric
displayed. Bottom line is you dont need to use validation. Im leaning towards make good market validation tests, then build systems. Then wf the best
ones, then chose one with good metrics, wf stability and validation on other time frames.
Soon we will have WF multiple time frames.
|
|
|
JasonT
Junior Member
Posts: 61
Registered: 6-6-2018
Member Is Offline
Mood: No Mood
|
|
Awesome thanks Peter
|
|
|
![[*]](images/davis/default_icon.gif){kind=icon}
