Machine learning in trading: theory, models, practice and algo-trading - page 373
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
Excerpt from Reshetov's article explaining how his RNN works.
"This article discusses in detail the problem of neural network retraining, identifies the causes of its occurrence, and proposes a way to solve this problem.
1. Why a neural network is overtrained?
What is the reason for retraining neural networks? Actually there may be several reasons for it:- The number of examples in the training sample is not enough to solve out-of-sample problems.
- The input data is unevenly distributed by the degree of correlation to the output data in different samples, which is very often the case when processing non-stationary data. For example, in a training sample, the correlation of any input parameter or several input parameters with respect to output values is significantly higher than in a test sample, or worse, correlation coefficients in different samples differ in sign. It is easy to check this by calculating correlation coefficients for all parameters in different samples before we train the neural network. And to get rid of this drawback is also quite simple, namely training examples decompose into samples randomly.
- Input parameters are not associated with output parameters, i.e. there is no cause-effect relationship between them - they are non-representative, and therefore there is nothing to train the neural network. And checking for correlations between input and output data will show a correlation close to zero. In this case you need to look for other input data, on which to train the neural network.
- The input data is highly correlated with each other. In this case you should leave the input data with the maximum correlation to the output data, removing the other data that correlate well with the remaining data.
All of the above causes of overtraining and methods of their elimination are common knowledge, since they have been previously described in various literature or articles devoted to neural network technology. "Only it is not a neural network in the full sense of the word, but a classifier ) That is why it is not retrained, but parameters are adjusted in the optimizer. Nothing prevents to use also neural network in optimizer, with different number of layers and different period of features and even their number, it would be even better
Read more here, there is a whole framework even https://www.mql5.com/ru/articles/3264
Bendat J., Pearsol A.
Applied Random Data Analysis: Translated from English: World, 1989.
At pp. 126
EXAMPLE 5.4. UNCORRELATED DEPENDENT RANDOM VARIABLES.
Two random variables X and Y are calledcorrelated if their correlation momentum (or correlation coefficient, which is the same thing) is different from zero; X and Y are called uncorrelated quantities if their correlation momentum is zero.
Two correlated quantities are also dependent. Indeed, assuming otherwise, we must conclude that µxy=0, which contradicts the condition, since
for the correlated quantities µxy ≠ 0.
The converse assumption does not always hold, that is, if two quantities are dependent, they can be either correlated or uncorrelated. In other words, the correlation moment of two dependent quantities may not be equal to zero, but it may also be equal to zero.
Thus, the correlation of two random variables implies their dependence, but correlation does not imply correlation. The independence of two quantities implies that they are uncorrelated, but the uncorrelatedness does not yet imply that the quantities are independent.
http://www.uchimatchast.ru/teory/stat/korell_zavis.php
1. no one is analyzing correlation - it's about the choice of predictors.
2. You repeated my point three pages earlier:"Dependence is a special case of correlation. If two variables are dependent, then there is definitely a correlation. If there is correlation, then there is not necessarily dependence."
3. cross-entropy just like correlation will not answer by the presence of functional dependence
That's where I was wrong - I admit it.
If random variables are independent, they are also uncorrelated, but you cannot infer independence from uncorrelation.
If two quantities are dependent, they can be both correlated and uncorrelated.
Only it is not a neural network in the full sense of the word, but a classifier) Therefore it is not retrained, and parameters are adjusted in optimizer. Nothing prevents to use also neural network in optimizer, with different number of layers and different period of features and even their number, it will be even better
read more here, there is a whole framework even https://www.mql5.com/ru/articles/3264
The above quote refers exactly to neural networks in the full sense of the word, and these are the problems he is trying to solve in his RNN
Yeah, and they are solved simply by enumerating all possible parameters and comparing them to forward... exactly the same thing can be done with NS. His RNN is retrained in the same way, we just choose the most optimal stable parameters comparing backtest with forward... everything is the same as with NS, only in case of NS we need to choose inputs-outputs in optimizer instead of weights.
If two quantities are dependent, they can be either correlated or uncorrelated.
Everything else is alternativeism and humanitarianism.
You finally got it))) Correlation only gives linear dependence and NS has nothing to do with it, also please do not confuse "correlation for non-linear regression" and "non-linear correlation", correlation is:
Everything else is alternative and humanitarian.
Again, forty-five....
What a strange person you are - above your post two posts it is written in black and white that the presence or absence of correlation does not mean there is any correlation at all and again correlation "gives" something to someone.
I'm sinking my hands....
Again, forty-five....
What a strange person you are - above your post two posts in black and white it is written that the presence or absence of correlation does not mean there is a correlation at all and again correlation "gives" something to someone.
I'm getting low....
You claimed that:
All MO is based on the fact that the input variables shouldcorrelate with the output variable.
Otherwise there is no point in ALL MO models.
In Data Mining, ALL VARIABLE VARIABLE SELECTION MODELS implement the mechanism of maximum correlation between the input variable and the output variable:
That is to say, they screwed up completely, they embarrassed themselves.
SZZ "The presence or absence of correlation doesn't mean correlation at all" - again - nonsense. Correlation does show the presence of a linear dependence, but there are also nonlinear ones that correlation does not show.
You stated that:
That is to say, they completely screwed up, embarrassed themselves.
SZZ "the presence or absence of correlation does not mean dependence at all" - again - nonsense. Correlation does show the presence of linear dependence, but there are also nonlinear correlations that correlation does not show.
I enjoy it when someone scientifically rolls someone :))
HI "The presence or absence of correlation does not mean dependence at all" - again - nonsense. Correlation does show the presence of a linear dependence, but there are also nonlinear ones that correlation does not show.
There is a classic example of a false correlation - the number of people drowning in US swimming pools is directly and strongly correlated with the number of movies Nicolas Cage was in.
There is a correlation - where is the GREAT correlation?