Machine learning in trading: theory, models, practice and algo-trading - page 3145

 
Maxim Dmitrievsky #:
I understand that, you could also look into causal forest. By the way, I haven't studied it, if someone will figure it out, it would be interesting to read about experiments with it
I don't understand Sanych's approach :) he is looking at the RMS error. Or RMS in a sliding window.

No. My sko is the "predictive ability" deviations. Nothing to do with the estimation of the model itself

 
СанСаныч Фоменко #:

No. My sco is about deviations in "predictive ability". It has nothing to do with the evaluation of the model itself

It's not just yours, but any MOSH person's :)
Cross-validation is common.
For some reason you just think you are doing something different.

If you estimate through MO, you will get comparable estimates. Because it works well, no worse than homemade estimates.

My conclusion is based on your description.
 

In the course of a similar experiment of selecting informative features, I tried all the ways. It's not difficult. Starting from correlation, mutual information and knn, through OLS and SVM to forest, bousting and neural networks (I didn't touch deep ones). It turned out to be best through bousting. OLS is in second place.

The reasoning is very simple: if bousting is cut down to one tree with one split, it is possible to evaluate mutual information, sample or permutation entropy and partly OLS.
 
СанСаныч Фоменко #:

No. My sco is about deviations in "predictive ability". Nothing to do with the evaluation of the model itself

Is it possible that the parameters of the model jump very much from step to step? That is, despite good "predictive power" at each step, the desired dependence is arranged very differently and is constantly changing. If so, this may well be a form of overtraining.

 
Maxim Dmitrievsky #:
I understand that, you could also look into causal forest. By the way, I haven't studied it, but if someone understands it, it would be interesting to read about experiments with it

It seems to be the same random forest, but with a causal interpretation. So you, as a populariser among us of forests and now of causal forests, have the cards in your hands).

Still, I don't understand the application of causal for trading yet. A quick googling did not help to find direct applications, only indirect ones - like studying the influence of stocks on Forex.

 
Aleksey Nikolayev #:

It seems to be the same random forest, but with a causal interpretation. So you, as a populariser of forests and causal interpretation among us, have the cards in your hands).

Still, I don't understand the application of causal for trading yet. A quick googling did not help to find direct applications, only indirect ones - like studying the influence of stocks on Forex.

It takes a lot of mental effort when dealing with the unknown :) there is no such thing on google, nor were there any clear general manuals until recently.
 
Aleksey Nikolayev #:

Is it possible that from step to step the parameters of the model jump very much? That is, despite good "predictability" at each step, the desired dependence is arranged very differently and is constantly changing. If so, this may well be a type of overtraining.

In my case it is impossible to answer your question: the model is being retrained at each step, and naturally the feature set may be different at different steps.

The classification error varies from 20% to 10%. 25% has never happened.

 
Maxim Dmitrievsky #:

In the course of a similar experiment of selecting informative features, I tried all the ways. It's not difficult. Starting from correlation, mutual information and knn, through OLS and SVM to forest, bousting and neural networks (I didn't touch deep ones). It turned out to be best through bousting. OLS is in second place.

The reasoning is very simple: if bousting is cut down to one tree with one split, it is possible to evaluate mutual information, sample or permutation entropy and partly OLS.

None of the above algorithms do NOT give predictive power, nor do hundreds of MO algorithms that stupidly calculate importance, which shows how often the algorithm uses a feature: If fed rubbish to an MO algorithm, any MO algorithm will compute the importance of that rubbish.

 
СанСаныч Фоменко #:

None of the above algorithms give predictive power, nor do hundreds of MO algorithms that stupidly calculate importance, which shows how often a feature is used by the algorithm: If you feed rubbish into an MO algorithm, any MO algorithm will calculate the importance of that rubbish.

The classification/regression error gives. I think enough of these strange games to play, you go round in circles :) And there is such a door to get out.
 
Maxim Dmitrievsky #:
Classification/regression error gives. I think enough of these weird games to play, going round in circles :) There's a door to get out.

We go round in circles in two because you are stupid, supposedly not understanding what I write.

There won't be a code. Think for yourself.