Machine learning in trading: theory, models, practice and algo-trading - page 364
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
Look here.
When evaluating the importance of predictors, it is important to remember that it is a complex quantity, not determined by information criteria alone. and
To my understanding, the importance of predictors, determined both by the means integrated into the basic algorithm and by separate means, does not solve the problem, because this importance is actually the frequency of using the predictor during the classification. If the predictors include Saturn's rings, coffee grounds , then the algorithm will use these predictors as well.
I have written many times on this thread and on other threads that predictors must "relate" to the target variable.
Let me give you an example again.
The target variable is male/female.
Predictor: clothing.
For countries with strict mores, if the predictor consists of pants and skirts, it splits into two non-overlapping subsets, each of which uniquely identifies the target variable, i.e. the classification error is zero.
If we start to include unisex, or Scottish style clothing in the predictor, then there is an intersection, which will determine the classification error. It is impossible to get rid of this error in principle.
It is not only my thoughts, but in this thread I gave links to publications with similar thoughts.
In my opinion, the importance of predictors, determined by both the means built into the basic algorithm and by separate means, does not solve the problem, because this importance is actually the frequency of using the predictor in the classification. If the predictors include Saturn's rings, coffee grounds , the algorithm will use those predictors as well.
I have written many times in this thread and in other threads that predictors must "relate" to the target variable.
Let me give you an example again.
The target variable is male/female
Predictor: clothing.
For countries with strict mores, if the predictor consists of pants and skirts, it splits into two non-overlapping subsets, each of which uniquely identifies the target variable, i.e. the classification error is zero.
If we start to include unisex, or Scottish style clothing in the predictor, then there is an intersection, which will determine the classification error. It is impossible to get rid of this error in principle.
These are not just my thoughts, but in this thread I have given links to publications with similar thoughts.
Truth is one, there are a million paths to it.
Everyone has the right to their own way. The main thing is that it doesn't lead to a dead end.
Good luck .
But I believe that it is possible to make a good self-optimizing device that will not work perfectly forever, but from time to time it will.
But this is clearly not on standard indicators and zigzag output will be :) It even sounds like some kind of child's play to me, just as an example if only.
I hope so too, but I'm far from illusions, which I had 3 years ago, that it can be done by configuring a dozen of parameters, popular library classifiers. It turned out to be very slow and trivial, especially in the evenings after work(((
ZZ is a bad target, I completely agree.
When assessing the importance of predictors, it is important to remember...
... the importance of predictors can be determined by their weights in the first layer
the algorithm for calculating importance in this function
Installed Open R, installed all the packages, VS 2017 still hangs when creating an R project
Python projects work fine
Took down R Studio and R 3.4, leaving only Open R, and it worked ) Apparently, they are conflicting.
And there is no point in R Studio, the same thing happens
Does it make sense to build a correlation of input variables to the target before training? And remove the most uncorrelated ones, to speed up calculations and increase the degree of learning?
Does it make sense to build a correlation of input variables to the target before training? And remove the most uncorrelated ones, to speed up calculations and increase the degree of learning?
We should not check for correlation of variables with the target (there should be one), but we should check for lack of correlation between variables - there should not be any (multicollinearity).
We should not check for correlation of variables with the target (there should be one), but we should check for lack of correlation between variables - there should not be any (multicollinearity).
I've already done the removal of the correlated inputs, I'm just wondering how else to improve the inputs.
So, I agree with you that there should be correlation with the target, that's why I want to additionally remove the most uncorrelated inputs with the target, for example with Kcorr<0.5 or 0.3. This should speed up the learning process without much degradation. But there is an assumption, that I will have to remove all inputs )))
On the used inputs (taken random from the technical indicators), so far I have not found any correlation with the target, the learning error = 0.44, i.e. almost a coin. Well, the balance is going down.