Machine learning in trading: theory, models, practice and algo-trading - page 3385

New comment
 
mytarmailS #:
I don't know, I don't really have anything to write.
I'll write how to break a wooden model into rules, so what?
Basically, my post already showed you everything.

Or are you referring to my old post? If so, in splitting I found no super healing properties, there are pluses that can not give the model.

1. You can drastically reduce the dimensionality of the model.



2. You can know the statistics of each rule (this is really important).

For example, we have a wooden model with 100 rules and we never know whether each rule worked once inside 100 rules (there is no pattern) or whether 10 rules worked 50 times (there is a pattern).
If we don't break the model, we won't know it and both models will be the same for us.

Well, in trees you can usually calculate the influence of each observation of each feature, its contribution to the model, for example through shap values. If we leave only useful ones and train something only on them, we will get an approximate analogue of rule search. With neurons, by the way, it is also possible.

It's hard to understand when only rules can be the only useful ones. Maybe for interpretability of the result. Although shap values also give good interpretability, sort of.
 
Maxim Dmitrievsky #:
Well, in trees you can usually calculate the influence of each observation of each trait, its contribution to the model, for example, through shap values. If you leave only useful ones and train something only on them, you will get an approximate analogue of rule search. With neurons, by the way, it is also possible.

It's hard to understand when only rules can be the only useful ones. Maybe for interpretability of the result. Although shap values also give good interpretability, sort of.
The influence of each trait, the influence of each observation and the influence of each rule are all different
 
mytarmailS #:
The impact of each feature, the impact of each observation, and the impact of each rule are all different
The rules are the elements of the model that link the features and labels. The only thing is that neural networks don't have discontinuity, but it can be artificially made.

What I'm saying is that I don't see much point in rules (smoking a pipe meaningfully).
 
Maxim Dmitrievsky #:
Rules are the elements of the model that link attributes and labels. The only thing is that neural networks do not have discontinuity, but it can be artificially made.

What I'm getting at is that I don't see much point in rules.

I'll try from Khabarovsk...


any model is a certain sum of patterns, exaggeratedly a pattern can be labelled as a TS.


Let's imagine that a model consists of 100 TCs.


It can be that in model #1 100 TSs made one deal.

It can be that in pattern #2 one TS made 100 trades, and the other 99 did not make a single trade.


how to calculate statistics for each TS?

If the model is a rule model, it is easy and clear.

If the model isneural?

 
mytarmailS #:

I'll try from near Khabarovsk.

If the model is neuron?

/\
!
!

Don't you and Sanych work in the same studio?
You choose examples that are well predicted by the NS. You train another NS on them only. Repeat several times, according to your taste. After a few rounds you'll have the NS with the best "rules".

Also easy and I wouldn't say incomprehensible.
 
Maxim Dmitrievsky #:

You choose examples that are well predicted by the NS. Train another NS on them only. Repeat several times, according to your taste. After several rounds, you will get the NS with the best "rules".
Well, we got a subsample on which neuronka predicts well. How do you know if it's one pattern in this subsample, two or twenty? You really don't know the difference?

 
mytarmailS #:
Well, we got a subsample where neuronics predicts well. How do you know if it's one pattern in that subsample, two or twenty? You really don't know the difference?

By the number of examples left. There are as many examples as there are patterns. It's an approximate rule, I'm not saying it's the same as a strict rule. But you can divide the sample further, up to a complete division for each pattern.
 
Maxim Dmitrievsky #:
By the number of examples left. There are as many examples as there are patterns.
There can be 200 examples and only 5 patterns.
An example is not a pattern, an example is one observation.
 
mytarmailS #:
There can be 200 examples and only 5 patterns.
An example is not a pattern, an example is one observation
If the error has already stopped falling or is equal to zero, you can divide the remaining examples into patterns by some measure of proximity :). Clustering, for example. And count how many are left. And even write an averaged condition for each pattern/cluster (take centroids of clusters), you will get an output rule.
 


mytarmailS #:

I'll try from Khabarovsk...


Any model is a certain sum of patterns, exaggeratedly, a pattern can be labelled as a TS.


Let's imagine that a model consists of 100 TS.


It can be that in model #1 100 TCs made one deal.

It can be that in model #2 one TS made 100 deals, and the other 99 did not make any deals.


how to calculate the statistics for each TS?

If the model is from the rules, it can be done easily and clearly.

If the model isneural?

The problem is not the number of times the model is used.

The problem is that the same model (tree?) on the same data predicts one label in some cases and a different label in other cases. This is what is called classification error. There are no predictors, at least with us, whose values can be categorised strictly into classes. and all the problems with leaves, trees and whatnot are derived from the values of the predictors.

1...337833793380338133823383338433853386338733883389339033913392...3533
New comment
Reason: