Machine learning in trading: theory, models, practice and algo-trading - page 3630
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
I wonder if anyone has ever been able to use standard approaches to feature selection to their advantage in trading.
I tried to use lasso and multicollinearity control to find FA signs - I did not find any significant ones at all.
I don't know about FA, I haven't tried to use it.
But about other features I can say that the picture lists secondary methods, and the main ones are not specified - it is calculation of "connection", "influence" of the feature on the target variable.
I have written about it many times for many years and cited a number of packages that allow to calculate such "links".
PS. "Connection", "influence" - this is NOT correlation and not the importance of the traits that machine learning algorithms produce.
Haven't done this in a while.
I think standard methods have been developed and tested on traits with a strong influence on the target. They are of little use in our field.
I came to brtforz, but not complete.
I train models on 1 trait (if 1000 traits - it means 1000 trainings), take the best one by balance, drawdown, etc. (it is possible directly on cross-validation). (you can measure it directly on cross-validation, I did it on valving-forward - i.e. test on gluing OOS pieces together).
Then to the best one I add 1 more and 999 training, then to a pair of the best one more and so on.
Approximately after 5-10 selected features the results of the model on OOS start to deteriorate. Up to 10 signs the time of such bruteforce is acceptable.
But the result is like everyone else's... As a result, since February I have been doing other things. There is no time for MO (.
But the result is like everyone else... I've been doing other things since February. I can't find time for the MO (.
And you test for what period? I teach for several years, half a year forward. From one to several months something there trades on the real.
Without super profits, I see )
In general, alpha is so low that even if you train the same model on the same data, it gives on the forward, it gives almost none. From a random number generator:)
You have to choose which one gives )
I think standard methods are developed and tested on features with a strong influence on the target. They are of little use in our field.
I've come to a brtforce, but not a complete one.
I train models on 1 trait (if 1000 traits - it means 1000 trainings), take the best one by balance, drawdown, etc. (it is possible directly on cross-validation). (you can measure it directly on cross-validation, I did it on valving-forward - i.e. test on gluing OOS pieces together).
Then to the best one I add 1 more and 999 training, then to a pair of the best one more and so on.
Approximately after 5-10 selected features the results of the model on OOS start to deteriorate. Up to 10 features the time of such bruteforce is acceptable.
What period do you test for? I teach for several years, half a year forward. From one to several months, something is traded on the real.
No super profits, I understand )
I tried different sized pieces of Walking Forward training, I stopped at 1-2 months. About 300 training for 5 years.
100-200% profit on OOS for 5 years, but with 2 big drawdowns up to 2 years.
That kind of thing is not interesting. I can make 100% in a day manually on martin, if the day turns out to be flat. And then drain on the trend))))))
And here 5 years for the same 100%, and for the next 5 years you can also lose if something fundamental shifts.... 2 days and 10 years with expectation of approximately the same result. What is better? Nothing.
About a year ago, and I threw in a chart.
I tried different sized chunks of Walking Forward training, stopped at 1-2 months. About 300 training for 5 years.
100-200% profit on OOS for 5 years, but with 2 big drawdowns up to 2 years.
That kind of thing is not interesting. I can make 100% in a day manually on martin, if the day turns out to be flat. And then drain on the trend))))))
And here 5 years for the same 100%, and for the next 5 years you can also lose if something fundamental shifts.... 2 days and 10 years with expectation of approximately the same result. What is better? Nothing.
What 5 years, 5 months maximum forward )) prepare so that the forward is less, but more trades
What 5 years, 5 months max forward )) prepare so that forward less, but deals more
Forward is not 1, but about 300 glued together. Valking-forward method. As in real life: trained, traded for a month, trained again and traded for another month, etc.
Forward is not 1, but about 300 glued together. Valking-forward method.
It's also from the evil one
If it's for beauty and theory, sure. But not on the FH.Don't know about FA, haven't tried to use it.
But about other attributes, I can say that the picture lists secondary methods, but the main ones are not specified - this is the calculation of "connection", "influence" of the attribute on the target variable.
I have written about it many times for many years and have given a number of packages that allow to calculate such "links".
PS. "Connection", "influence" are NOT correlation and not the importance of features that machine learning algorithms produce.
From my scrappy recollection, your approach can be attributed to the Filters section from the image I posted. There's more than just correlation there. And there is no attachment to any MO model. Probably not all specific methods are given yet (for brevity).