Machine learning in trading: theory, models, practice and algo-trading - page 519
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
So, what's up? Any helpers?????
What topic???? Buddy... Who are you?
Just maybe you're new here and don't know me. I'm kind of an AI guy in general,too..... What are you... You're from around here. :-)
Although I'm kind in principle and everything you do here will work ONLY when the data will be the reason for the price. Then any TS will work fine, even if it's 1 bar or 15 bars ahead of the forecast (of course 15 bars will be worse than 1, but not the point). Not the point... The point... RTS index which has OI. As the meaning of volume.... And the problem is SOLVED. ANYTHING, even a prediction, even a classification.
And what did you want to say with your phrase, dear.....
:D
Now I want to say - fuck you. My dear man. Thank you.
Exactly, neither last time nor this time you didn't say anything, as if you farted into nothing...... We don't need that kind of interlocutor. Sorry....
Paradoxically, classification is the same as regression.
Only for regression the output is a real, and for classification it is a probability.
Regression targets are continuous curves, while classification targets are (0, 1) or (-1,+1) impulses.
This output is then translated into the corresponding class (ifelse(y > 0.5, 1, 0).
If there are three classes - -1,0,1 (buy, sell, wait). Is it via 0.5 (k=0;if(y<-0.5)k=-1; if(y>0.5)k=1; the middle point between classes, but the probability of class 0 will be twice higher than 1 or -1) or via 0.33 (k=0;if(y<-0.33)k=-1; if(y>0.33)k=1; so all classes have equal probability)?
A good way to classify is as follows. A separate output is allocated for each class. The cross entropy learning loss function is used. When teaching 1 you can assign only one class. It is necessary to provide for the class of no signal. For example buy, sell, and do nothing. These are separate classes. Feeding all values to one output is inefficient because one neuron may not learn to divide, say, 10 classes.
The derivatives show the direction of the trend. The derivatives of 2 MAs and the difference between them completely describe the state of the system. You asked the thread yourself.) The NS will figure it out from there.
However, it is up to your business.)
the derivatives of the 2 mashes and their difference depend only on 4 bars, they cannot describe the state of the system well just in any way
1) Is the transfer to a class always done through the middle of the range?
2) And if 3 classes -1,0,1 (sell, wait, buy). Transfer through 0.5 (k=0;if(y<-0.5)k=-1; if(y>0.5)k=1;middle between classes, but the probability of class 0 will be twice as high as 1 or -1) or through 0.33 (k=0;if(y<-0.33)k=-1; if(y>0.33)k=1;so all classes have equal probabilities)
In R it is usually possible for the classification to have variants of the results:
You set the mode "class probability" and calculate the class on your own, for example in half or 30/70. Or you can do this: less than 30% one class and more than 70% another class, and the gap between them as NA.
1) Is transferring to a class always done through the middle of the range?
2) And if there are 3 classes -1,0,1 (sell, wait, buy). Transfer through 0.5 (k=0;if(y<-0.5)k=-1; if(y>0.5)k=1;middle between classes, but the probability of class 0 will be twice as high as 1 or -1) or through 0.33 (k=0;if(y<-0.33)k=-1; if(y>0.33)k=1;so all classes have equal probabilities)?
Talk about two classes. Continuous classifier output can be translated into class labels using a threshold:
Two variants of the result are possible:
As a rule, the classifiers have a continuous numeric variable (degree of support) as an output. Degrees of support for a given input X can be interpreted in different ways, the two most common being validity in proposed labels and estimation of possible probabilities for classes. Models give out class probabilities poorly, so they usually need to be calibrated. See calibrate::CORElearn. The outputs after softmax activation functions are closest to class probabilities.
Good luck