Machine learning in trading: theory, models, practice and algo-trading - page 632
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
I'm interested in reinforcing lerning, so I found an interesting article, and I'm trying to buy it and add it to the bot
https://hackernoon.com/the-self-learning-quant-d3329fcc9915
Only it has some kind of dope, at the end only learns to buy for some reason
I have the same situation - the NS classifier often falls into one class. Classes need to be aligned. But I have 95% of one class and 5% of the second. *20 volume of lines is not desirable.
This is about the same as in that article and I get.
I think I'll switch to regression/forecast to experiment and see what will happen...
I'm interested in reinforcing lerning, so I found an interesting article, I'm trying to buy it and add it to the bot
https://hackernoon.com/the-self-learning-quant-d3329fcc9915
I tried to buy a bot but it had some strange idea that he is only learning to buy at the end
The article is useful in that he links to two good libraries in the code.
One for loading different market data, the other for applying on data different indicators. Took it for myself.
https://www.quandl.com/tools/python
https://mrjbq7.github.io/ta-lib/
I have the same situation - the NS classifier often falls into one class. Classes need to be aligned. But I have 95% of one class and 5% of the other. *20 volume of rows is not desirable.
This is about the same as in that article and I get.
I'm thinking of switching to regression/forecast to experiment and see what happens...
I also encounter this situation.
Feature selection, regularization, and lots of dropouts make the model more stable. All without alignment, pure time series flow.
I've been messing around with classes a lot for some reason. In my case (ml-assistant), signal persistence happens with a single number [-1,1]. That is, in theory, you can immediately build a regression (which also comes out quite well), but the prediction is less "readable" there.
To classify, you have to break them into classes, and the softmax idea is to equate the sum of classes to 1.0. But in the end it turned out to be most correct just to divide them relative to zero, the model itself normalizes the sum of classes to 1, regardless of the targeting.
But when I add a third pass class, for some reason I keep overtraining. Maybe I'm not preparing it correctly. =)
I have the same situation - the NS classifier often gets bogged down in one class. Classes need to be equalized. But I have 95% of one class and 5% of the other. *20 volume of rows is not desirable.
This is about the same as in that article and I get.
I'm thinking of switching to regression/forecast to experiment and see what happens...
Out of a 10-12k sample: ~800 are class 1, the rest are class 0. That's the reality).
I read in a smart book that the quantitative ratio of classes in the training sequence should be close to reality.
Out of a sample of 10,000-12,000: ~800 are class 1, the rest are class 0. This is the reality).
In a clever book I read that the quantitative ratio of classes in the training sequence should be close to reality.
and other smart books say that the classes should be balanced/combined.
i.e. in forex, we have no idea about the population, and according to sound logic, the number of elements in classes should be approximately equal.
Just an interesting video about randomness from a smart guy
Hello all!!! Can someone explain the meaning of negative entropy? In simple terms......
Here is what clever people write:
"Non-entropy θ corresponds to some "structural" value. which is determined by the characteristic amount of past memory. At large θ complex hierarchical structures arise in a wide range, at small θ - structures in a small range, and at θ → 0, i.e. absence of information about the past, there is a marginal transition to Markov processes".
In fact, this is the value that characterizes the non-markovianity of the process. Since I work with probability distributions, I found out that this thing is responsible for the type of "tails" of distributions.Hello all!!! Can someone explain the meaning of negative entropy? In simple terms......
Entropy is a measure of disorder or chaos, non-entropy (negative entropy) is a measure of decency and degree of self-organization. Undissolved sugar in a glass system has minimum entropy, after complete dissolution - the system has maximum entropy. To return the system to its initial state, it is necessary to import entropy from outside - to create a crystallization condition. As applied to NS, the system must constantly absorb new information from outside and get rid of unnecessary information, otherwise, increase of entropy will lead to its death.
Entropy is a measure of disorder or chaos, non-entropy (negative entropy) is a measure of orderliness and degree of self-organization. Undissolved sugar in a glass system has minimum entropy, after complete dissolution - the system has maximum entropy. To return the system to its initial state, it is necessary to import entropy from outside - to create a crystallization condition. In the case of NS, the system must constantly absorb new information from the outside and get rid of unnecessary information, otherwise, an increase in entropy will lead to its death.
Excellent comment. Exactly so.