Machine learning in trading: theory, models, practice and algo-trading - page 879

 
Yuriy Asaulenko:

I believe - the only one not for today, but in general the only one for MO and HC in configurations of reasonable complexity. First of all, we limit the fields of application of NS and MO, and already on this we apply NS and MO.

And solving problems like "in general, all at once" - that's for AI).

Rather, NS is a trainable logic of decision making. Originally it was supposed to replace it in standard strategies, so that I wouldn't bother with its writing.

Well, "never in general" is too strong a statement. I hope for the appearance of a sensible quantum computer in the next 50 years))) And then maybe a real AI.)

But yes - current development of processors has reached the physical limit of placement on the plane. And in three dimensions it is much more difficult to do everything.

 
elibrarius:

Targeted - you have regression, not classification. I've given up on regression for now. I think it's better to train 2 neuroschemes, by number of targets, but I haven't done enough experiments with regression myself - experiment on your own.
The sequence of columns isn't important, the main thing is to tell NS that these are targets. Row sequence is probably better to have freshest data at the end (but not necessarily), many packages shuffle all rows by default for even training. Otherwise NS may deadlock somewhere in the middle (local minimum) and not get to fresh data. Fresh data (last 10-20%) can be 2 - 3 times fed so that the network better learns the latest market trends - also an opinion that I haven't tested in practice.
Check out the blog of the topicstarter - he taught regression there, a lot of good thoughts. But in the end he wrote that he had found some error in the code that made all results invalid.

So there are no clear and unambiguous answers, that's why all are silent)

Just what I want to convert the target into a classification! My understanding is that the tree only eats 0 and 1 (i.e. two logical values), and in my case the buy and sell results need to be separated separately and trained(classified?) as separate trees.

Thanks for the reply! I rejected regression, because its absolute values will be noise in a non-stationary system, and will use regression as a logical solution, for example to find prices relative to the channel.

About mixing data is clear, but strange, and I understood that if we want to inform NS/Decision Tree about past events, we should make a copy of predictors with a shift, increasing the depth for each shift by an order of magnitude?

 
Aleksey Vyazmikin:

About mixing the data is clear, but strange, and I understood that if we want to inform the NS/Decision Tree about past events, we have to make a copy of predictors with a shift, increasing the depth for each shift by an order of magnitude?

Why else would you make a copy? Each row in your table is the same predictors going back in time.

 
Maxim Dmitrievsky:

What kind of questions?

You cannot teach a neural network the expected profit, you need to teach the things that the neural network should classify or approximate, i.e. specific buy/sell actions or specific market conditions

As for predictors, which are better and which are not, no one will give you the answer, because you need to study them and see how the TS works, that's what this whole topic is about, for the most part, but everyone has different ones

about neural networks - you can use alglib or R or Python - it's been written about 100 times

about the training sequence - it doesn't matter which data is newer and which is older, you need to understand how the NS works - it minimizes the error for all cases in the aggregate

You can find a lot of basic information about neural networks on youtube and google, it's not discussed here because there's no point

I wrote - it is necessary to study everything from the basics, then there will be an understanding of what and where, otherwise it's an empty job.

Six months ago I myself still do not understand them, now more or less. Re-read a ton of literature and hundreds of hours of videos, then began a kind of synthesis of knowledge. And yes, my questions, too, very few people had clear answers, I investigated myself :)

And your questions are too vague, it turns out that you need to be a psychic to start guessing, because you need to observe a huge number of details that would make something work, rather than look at the table and understand everything :))

My contribution to the development of MO on the resource, I have made in the form of two articles, whether I will continue to write it until I do not know, because there has already begun the area of sacred grails (just kidding)

Why can't I teach entry points where the maximum profit to buy or sell is expected?

I have enough predictors - I am slowly incorporating them into my script. I do not know yet if they are good or bad in the end...

About neural networks - I was interested in specific names of NS, which work better with target ones in form of classification... And so far I haven't been able to set up any normal package yet :( That's why I'm testing with "Deductor Studio Academic" - very understandable interface, everything is in Russian, there are trees and neural network, it looks good for a beginner, but the minus is that you can't export the results.

Here I do not know whether it is a good result for the tree? I trained 50% of sample and tested it on 50%.


I'm studying the basics, re-read some articles here and watch lectures on NS, but not everything is clear there, and I have no one to ask...

Thanks for your attention.

 
elibrarius:

Why else do you need a copy? Each row in your table is the same predictors going back in time.

Then how is it possible to mix them? After all, their sequence is broken...

Does the tree look at the sequence?

 
Aleksey Vyazmikin:

Then how is it possible to mix them? After all, their sequence is broken...

Does the tree look at the sequence?

Violation is the point, to avoid getting into one of the local minima. However, mixing is not necessary - you can learn without it; depending on the presence of local minima in your data, and the presence of other methods for skipping local minima, if they do exist.

I don't do trees.

 
Aleksey Vyazmikin:

Why can't I teach entry points where the maximum profit is expected for buying or selling?

I have enough predictors - I am slowly putting them into the script. So far, however, it is not clear whether they are good or bad in the end...

About neural networks - I was interested in specific names of NS, which work better with target ones in form of classification... And so far I haven't been able to set up any normal package yet :( That's why I'm testing with "Deductor Studio Academic" - very understandable interface, everything is in Russian, there are trees and neural network, it looks good for a beginner, but the minus is that you can't export the results.

Here I do not know whether it is a good result for the tree? I trained 50% of sample and tested it on 50%.


I'm studying the basics, re-read some articles here and watch lectures on NS, but not everything is clear there, and I have no one to ask...

Thanks for your attention.

Too good - an error of less than 10%. Don't your predictors peek into the future? That's usually why the error is so small. Or target the past? Zigzags, for example? Or you predict a 0 bar and the predictors are built with Close 0 bar.
 
elibrarius:
It makes sense to avoid hitting one of the local lows. However, mixing is not necessary - it is possible to learn without it; depending on the presence of local minima in your data, and the availability of other methods to skip lock minima, if they do exist.

I don't do trees.

Well, how should the sequence (chronology) of data be presented in the file - in the beginning is the newest (2018) or the oldest (2017)?

 
Aleksey Vyazmikin:

Okay, and the sequence (chronology) of data should be in the file as presented - in the beginning the newest (2018) or vice versa the oldest (2017)?

NS usually processes the data from the first lines - i.e. the oldest data should be in the first lines, and fresh data should be at the end, so that the last steps of training were done on them.
 
elibrarius:
Too Good - less than 10% error. Don't your predictors peek into the future? That's usually why the error is so small.

Not as far as I know - yesterday I found one predictor that uses 1 bar, but my target does not depend on the number of bars since the position was opened (i.e. the dependence is not set), I closed using stop loss, which works on the indicator.

All predictors work at bar opening - I do not even know how to detect those that peek - their importance should be high, according to the idea?

I do not see it on the picture...