Machine learning in trading: theory, models, practice and algo-trading - page 2532

New comment
 
You don't really need a reference one. It is supposed to learn on those quotations on which you plan to trade. That's why the spread of a particular brokerage company is needed.
 
elibrarius #:
The benchmark is not really needed. It is supposed to learn on those quotations on which you are going to trade. That's why the spread of a certain brokerage company is needed.
If you want to check the real performance and start relying on it, what's the sense of looking at the spread that was several years ago?
 
Maxim Dmitrievsky #:
Then look at the real performance and then dance from that, what's the point of looking at the spread that was a few years ago?

The model must adjust to the current situation with regular retraining. 5 years ago with its spread, and current data with its spread.

 
elibrarius #:

A model with regular retraining must adjust to the current situation. 5 years ago with its spread, and current data with its spread.

Sounds questionable, extra degree of freedom, given imperfect history

 
Maxim Dmitrievsky #:

Sounds dubious, an additional degree of freedom, given the imperfection of history

Better an imperfect history from the current broker than a perfect one from another one.
 

Have you tried the opposite, adding artificial noise to distort a series of observations randomly?

p/s yahho is often scolded for some reason, advising to take data from paid suppliers

 
LenaTrap #:

Have you tried the opposite, adding artificial noise to distort a series of observations randomly?

p/s yahho is very often scolded for some reason, advising to take data from paid suppliers

Lena?!

Tell me more.

This approach is very interesting.

 
The main thing is not to forget to remove the noise from the equity back.
 
elibrarius #:
Better an imperfect history from a current broker than a perfect one from another.
Worse, there may be garbage out there, better to educate yourself on proven sources, otherwise you can get caught up in false dependencies and other inadequacies. And get duped.
 
LenaTrap #:

Have you tried the opposite, adding artificial noise to distort a series of observations randomly?

p/s yahho is very often scolded for some reason, advising to take data from paid suppliers

Sometimes helps from overtraining a little, equal to dropout
1...252525262527252825292530253125322533253425352536253725382539...3516
New comment
Reason: