Machine learning in trading: theory, models, practice and algo-trading - page 2493

[JeeyCi]

eccocom #:
Read the TensorFlow documentation, everything is a constructor... practically. Really it's black boxes. If it's interesting I can give you the code of the perceptron written manually, and by the way it's all matrix calculations, that's what it's all built on

I just came back to Evgeny Dyuka's post and his ServerNN (to read the code) on TensorFlow

I just came across a kind of critique of Python libraries on the hobby

The sklearn stack, xgboost, lightGBM packages had built-in methods to estimate feature importance for "wood models":

1.Gain
This measure shows the relative contribution of each feature to the model. To calculate this, we go to each tree node, look at which feature leads to a node split, and how much the model uncertainty is reduced according to the metric (Gini impurity, information gain).
For each feature its contribution is summed over all trees.

2.Cover
Shows the number of observations for each feature. For example, you have 4 features, 3 trees. Suppose chip 1 has 10, 5, and 2 observations in trees 1, 2, and 3 respectively. Then the importance of this chip is 17 (10 + 5 + 2).

3.Frequency
Shows how often a given phicha occurs in tree nodes, that is, the total number of tree splits into nodes for each phicha in each tree is counted.

The main problem with all these approaches is that it is not clear how exactly a given feature affects model prediction. For example, we have learned that income level is important for estimating the bank customer's ability to pay the loan. But how exactly? How much does higher income shift model predictions?

Here, however, about "wooden models" - I'd like to avoid them, because they eat a lot of RAM (and I'm a fan of speed and compactness)...

- anyway, the answer to your question is CONVENIENTLY, INTERESTING... yes, that's what i thought, Neural Networks is all matrix calculus (that's why i was asking about it and libraries for it)

p.s.

I just don't understand, if those libraries in the quote can't predict the coefficients of a polynomial (answer: by how much), how will they describe a dynamic model? (not linear, not regression [which is the same thing])... that's why so far I've wondered about making a model (trying to remember diferentiation... with difficulty)

[eccocom]

JeeyCi #:

I just came back to Evgeny Dyuka's post

I just came across some criticism of Python libraries on the hobby

The sklearn stack, xgboost, lightGBM packages had built-in methods to estimate feature importance for "wood models":

1. Gain
   This measure shows the relative contribution of each feature to the model. To calculate this, we go to each tree node, look at which feature leads to a node split, and how much the model's uncertainty is reduced according to the metric (Gini impurity, information gain).
   For each feature its contribution is summed up for all trees.
   
2. Cover
   Shows the number of observations for each feature. For example, you have 4 features, 3 trees. Suppose chip 1 has 10, 5 and 2 observations in trees 1, 2 and 3 respectively. Then the importance for this chip is 17 (10 + 5 + 2).
   
3. Frequency
   Shows how often a given phicha occurs in tree nodes, i.e. the total number of tree splits for each phicha in each tree is counted.
   

The main problem with all these approaches is that it is not clear how exactly a given feature affects model prediction. For example, we have learned that income level is important in estimating the bank customer's ability to pay the loan. But how exactly? How much does higher income shift the model predictions?

Well, for my taste Python is criticized by those who don't know how to make it. Everything else is easier to do not taking ready made solutions and write them using libraries, and by the way using for example numpy you will probably get more speed than C++ as it is written in Fortran. The more so now Python is easily integrated with MT5 so to say in straight line, it allows to use AI without any servers with sockets.

[Forester]

JeeyCi #:

I just came back to Evgeny Dyuka's post

because I came across a kind of criticism of Python libraries on the hobby

The sklearn stack, xgboost, lightGBM packages had built-in methods to estimate feature importance for "wood models":

1. Gain
   This measure shows the relative contribution of each feature to the model. To calculate this, we go to each tree node, look at which feature leads to a node split, and how much the model's uncertainty is reduced according to the metric (Gini impurity, information gain).
   For each feature its contribution is summed up for all trees.
   
2. Cover
   Shows the number of observations for each feature. For example, you have 4 features, 3 trees. Suppose chip 1 has 10, 5 and 2 observations in trees 1, 2 and 3 respectively. Then the importance for this chip is 17 (10 + 5 + 2).
   
3. Frequency
   Shows how often a given phicha occurs in tree nodes, i.e. the total number of tree splits for each phicha in each tree is counted.
   

The main problem with all these approaches is that it is not clear how exactly a given feature affects model prediction. For example, we have learned that income level is important in estimating the bank customer's ability to pay the loan. But how exactly? How much does higher income shift model predictions?

All of these methods are unworkable. Here's checking them out https://www.mql5.com/ru/blogs/post/737458
Trees and forests count quickly - you can also just add/remove features one at a time. And find the ones that improve and the ones that worsen the result.

[eccocom]

As for models, it is not about them, but the fact that AI is essentially an approximator and that is its strength in all other cases except the market, because there are too many fish and gozils under the smooth sea, which cause price waves. While the postulate that everything is built into the price is true, there is too much of it. The neural network can only be a small part of some system, and who does not agree with this is engaged in self-deception.

[mytarmailS]

eccocom #:
As for models, it's not about them, it's about the fact that AI is essentially an approximator and that's its strength in all other cases except the market, because there are too many fish and gozils under the smooth sea that cause price waves. While the postulate that everything is built into the price is true, there is too much of it. A neural network can only be a small part of some system, and whoever disagrees with this is engaging in self-deception.

And the one who calls approximators AI, what does he do? )

[eccocom]

mytarmailS #:
And whoever calls the AI approximators what do they do? )

Well, I was probably wrong to call it AI, because I meant just a neural network, and it's a special case of AI. I'm doing the same as anyone else, searching for the Grail, doing it optionally, without fanaticism.) I came here to find ideas for models (in the sense of a neural network), there is something to share myself. 20 years of experience is certainly not available, but for a hundred variants of the NS I have gone through. Maximum efficiency at short distances, which reached about 68%, but you probably understand that it is not stable and situational.

[mytarmailS]

eccocom #:
Well, I was probably wrong to call it AI, because I meant just a neural network, and this is a special case of AI. I'm doing the same as everyone else, searching for the Grail, doing it facultatively, without fanaticism). Suda came to find ideas for models (in the sense of neural network), there is something to share myself. 20 years of experience is certainly not available, but for a hundred variants of the NS I have gone through. Maximum efficiency at short distances, which I reached about 68%, but you probably understand that it is not stable and situational.

Yes, of course I understand...

But beware of experts with a 20-year experience, which nag the code from others.)

[eccocom]

mytarmailS #:
Yes, of course I understand...

But beware of experts with 20 years of experience who are begging for code from others. ;)

I don't mind the code, if a person is good). By the way, I read the article by Mihail Marchukajtes. I use some ideas that were used there too, but I had completely different reasons than the author did. It would be interesting to discuss which model variants with which layers turned out to be the most effective. Mikhail, judging by circumstantial evidence, seems to use Sequential() full-linked layers.

[mytarmailS]

eccocom #:
I don't mind the code, if the person is good). By the way, I have read the article by Mihail Marchukajtes, some of the ideas that are used there I also use, but my reasons were completely different from those of the author. It would be interesting to discuss what variants of models with what layers turned out to be the most effective. Mikhail, judging by indirect evidence, seems to use full-link Sequential() layers.

He doesn't use neurons at all, he uses a ready-made author's mediocre program (some hybrid of SVM and MGUA), he programs neither in Python nor in R, so the word Sequential() is unfamiliar to him, he has 20 years of experience in neural networks, it's too early to get interested in such things...

[eccocom]

mytarmailS #:
He doesn't use neurons at all, he uses a ready-made author's mediocre program (there is some hybrid of SVM and MGUA), he doesn't program in Python or R-code, therefore the word Sequental() is not familiar to him, 20 years of experience in neural networks, it is still early to get interested in such things...

)))). I see, it seems that you have a solid baggage of communication. I myself probably write 8 out of 10 Hindu code)), but I'll be glad if I can share my experience all the same. And it would be great to find those who write in Python, and not only the neural network, but the whole structure of EAs in general.