Machine learning in trading: theory, models, practice and algo-trading - page 1615

New comment
 
Aleksey Vyazmikin:

It's not about shrinkage, it's about statistical behavior of the predictor on the sample outside of the split - this should reduce the randomness of selecting the predictor value.

By the way, does AlgLib do the grid on every split or once and then use this grid? As I understand it, the developers of CatBoost say that the grid is made once at them.

There is no randomness. The best available partition of each predictor is selected. Randomness is in the forest, when each tree is fed not all predictors, but for example half of the randomly chosen ones.

It learns once. There is no retraining. For trees/forests seem to have no re-learning at all, apparently from the fact that to re-learn quickly enough.
And why the grid? In trees, knots and leaves.

 
elibrarius:
By the way, what I don't like about boostings is that the recommended tree depth is 7-10.
That is, if we have 100 predictors, and the division there also starts in the middle of each predictor. Then with a high probability we will have 7 different predictors divided in the middle. Maybe 1 or 2 will divide to a quarter, hardly any smaller.
Or in boosting algorithms, the algorithm doesn't work by half division, but in smaller chunks? Does anyone know?
And who uses what tree depth?


 
Maxim Dmitrievsky:


Forty-seven minutes is a pity... to listen to the basics, which are mostly known. Only one specific question is of interest. If you know, tell me)

 
elibrarius:

Forty-seven minutes is a pity... to listen to the basics, which are mostly known. Only one specific question is of interest. If you know, tell me.

they are all built differently, you have to read the help for each one.

it all doesn't matter, if there are informative chips that are relevant to the target, then any method works

I was comparing forest with boosting on similar features. Boosting has less overfit, overall +-

 
Maxim Dmitrievsky:

they are all built differently, you need to read the help for each

it all doesn't matter, if there are informative features that are relevant to the target, then any method works

I was comparing forest with boosting on similar features. Boosting has less overfit, in general +-

How deep did you set the boosting?
 
Maxim Dmitrievsky:

they are all built differently, you need to read the help for each

it all doesn't matter, if there are informative features that are relevant to the target, then any method works

I was comparing forest with boosting on similar features. Boosting has less overfit, in general +-

That's what I mean. If the inputs are suitable for the target, then any method will work. What I wanted to check because I have really good inputs. But only Reshetova optimizer proves it, and as you understand one expert is not enough to make a subjective assessment. It's not about the method, each of the methods only requires a certain number of features (Max, because of you first used this word) someone needs a lot of them, someone a little, but in general, if the input data explain the output variable, then any method will work. And I think the results will be similar. It is a pity I cannot use my inputs in other systems than Reshetov. It is necessary to go to the forum of programmers, where I have to go :-(.
 
Elibrarius:
What depth did you set for the boosting?

2 to 10, the greater the depth, the greater the fit

optimally 3-7

gradient step can also be changed. In general, it does not matter, the results are less scatter, less offset, less signals, etc... And the average picture is preserved. This is a matter of optimization, it has nothing to do with the quality of the feature.


 
Max, honestly, thank you for the video about the natural neuron, but this video is not so good. The fact is that I have a theory of retraining about which I think for a long time and have built it to me quite adequately. I'm sure Yandex employees will be interested to hear it. Eh... I wish I could find the strength to record a video. I'm always drunk, then funny. It's not like that :-(
 
Mihail Marchukajtes:
Max, I want to thank you for the video about natural neuron, but this video is not so good. The thing is that I have a theory of retraining about which I've been thinking for a long time and built it quite adequately for me. I'm sure Yandex employees will be interested to hear it. Eh... I wish I could find the strength to record a video. I'm always drunk, then funny. It's not like that :-(

))) regularities should be sought through statanalysis, not by torturing neurons

For example, in my penultimate article I gave the EURUSD seasonal fluctuations for 10 years, by months. This year, everything is repeating itself. April-May will be the most interesting (from the nearest)
 
Maxim Dmitrievsky:

))) regularities should be sought through statanalysis, not by torturing neurons

That's what I mean. Before bothering JPrediction I leave out of 6000 000 bars only 150 pieces, which are statistically significant, and only then I look for that notorious law describing the output. The number of columns should be twice as big as the number of table rows, theoretically speaking, so the algorithm has a lot to choose from. As a result, the optimizer leaves 5 to 10 pieces out of 150 offered by me columns to form the final model.
1...160816091610161116121613161416151616161716181619162016211622...3476
New comment
Reason: