Machine learning in trading: theory, models, practice and algo-trading - page 3144

New comment
 
Lorarica #:

A published signal, very much such a concept.

Could have been a guy sitting there tapping his hands on the buttons.

It's been out of service for over two years. Isn't that a long time?

P.Z.

Try to understand by reading slower what you wrote above?

Is there at least one correct statement?

didn't read carefully

 

When will you realise that you need an education to work in the market.

And not just any education.

P.Z.

As you have here, everything is complicated.

For example, who will read 31400 messages?

Take it, cut it down, summarise it.

You've got smart people who know about this machine learning stuff, right?

What's wrong with my advice?

All the best, just for you.

P.Z.

 
mytarmailS #:

didn't read carefully

don't feed the trolls - otherwise they will flood the thread completely

 
Lorarica #:

Perdon.

7 years old branch, 31400 posts, where is the result?

And what happens if you delete 30000 posts if there is nothing in them?

Or is there? Who knows?

P.Z.

Banan/Anan=1.23

This is a communication thread. The topic of communication is MO. Here people communicate, share their impressions and sometimes intermediate results.

Nobody obliged her to gradually build a working trading advisor. So you can easily be written down as a baiter, be careful

 
СанСаныч Фоменко #:

Read the texts in Chinese and then retell the information from them.

Well, yes, I am a bad receiver of Chinese texts) That's what I wrote about. But this does not mean that there is no information in a particular Chinese text, because you can introduce an additional information converter - a translator from Chinese to Russian and then it will become clear whether there is information in a given Chinese text or it is just a random set of characters.

 

Regarding stability (c) SanSanych. If you add time to a set of attributes, you can compare its significance with others. If a trait is more significant than time, then it is stable. Perhaps this makes some sense)

For example, if one builds a decisive tree, only do so up to the first split in time. If the tree turns out to be empty, then all signs are bad. Some justification for this approach (for the case of trees) may be the similarity of split point search algorithms with change point detection of a time series. In both cases, the split of a single sample into two maximally different subsamples is usually sought.

 
I apologise profusely, but why can't fiches simply be tested through the model on new data in such a case?) stability is stability in Africa.

It has exactly the same orthodox log-loss at the window as mutual information or other similar metrics.

And in terms of efficiency it will be about the same, infinity multiplied by infinity, since 2 random series are compared.

(с)
 
Maxim Dmitrievsky #:
I apologise wildly, but why can't you just check the chips through the model on new data in this case?) stability is stability.
.

It has exactly the same orthodox log-loss on its window as mutual information or other similar metrics.

And in terms of efficiency it will be about the same, infinity multiplied by infinity, since 2 random series are compared.

(с)

It's hard to say. IMHO, a large window, several times larger than usual, is taken for analysis. Then we build a decision tree on it, adding time as a feature. If everything starts with time splits, we call the other signs bad, unstable. Even if these signs suddenly work well on smaller windows, there will still be instability, because the dependencies on different windows will be very different.

 
Aleksey Nikolayev #:

It's hard to say. IMHO, a large window, several times larger than usual, is taken for analysis. Then we build a decision tree on it, adding time as a feature. If everything starts with time splits, we call the other signs bad, unstable. Even if these signs on smaller windows suddenly work well, there will still be instability, because the dependencies on different windows will be very different.

I understand that, you can also look at causal forest. By the way, I haven't studied it, if someone will figure it out, it would be interesting to read about experiments with it
I don't understand Sanych's approach :) he is looking at the RMS error. Or RMS in a sliding window.
 
Aleksey Nikolayev point search algorithms with change point detection of a time series. In both cases, the split of a single sample into two maximally different subsamples is usually sought.

Adding. Blank, time is virtually a null feature

1...313731383139314031413142314331443145314631473148314931503151...3619
New comment