Machine learning in trading: theory, models, practice and algo-trading - page 232

 
It'snot:

I agree, in my opinion, if a person can at least just run the data and get the logos below 0.69300 (random) then has the right to talk about AI and MO here, the rest are not profitable

my result https://numer.ai/ai/toxic

https://numer.ai/ai/dr_tr

0.69184 got it right :)

Interesting contest, and interesting prizes.

And I also wonder what principle they use to create predictors. I usually have dependencies that fade with time, i.e. I trained the model on the first half of the file, and expected that on the second half of the file the error would grow with time, as usual. But there - no, everything is so stationary, without degradation over time. Although they promised trade-related data. Unusual. Perhaps the rows in the table are randomly shuffled.

 
Vladimir Perervenko:

This thread is getting too big and unreadable. I propose to start a new branch "RUserGroup" in which only specific issues of applying machine learning models in MT4/5 terminal in languages which allow to do this without problems will be discussed. I know two (R, Python). Discussions to be held with code provided. Experts with some experience in other languages are also welcome.

We can start with an example of convolutional network in previous posts.

Good luck

I am in favor of it! But they will shit all the same(!)

I have another question: what is the real advantage of a convolutional network over a conventional one, in your opinion? on the market, of course)

 
ivanivan_11:
You have an idea, you need to check it in 5 minutes (a day, a week). and it is unknown - is your idea viable or initially stillborn. what option will you choose - to use a ready-made package, which a student can figure out in 5 minutes, or write a complete infrastructure again yourself? if the second option, it is sadomasochism, and normal people with you not the way))
What infrastructure are we talking about? Are we talking about their methods and classes as the great Nikolai Kositsyn did? If so, it is more than normal, but not everyone is given.
 

I agree, in my opinion, if a person can at least just run the data and get the logos below 0.69300 (random) then has the right to talk about AI and MO here, the rest are not profitable

My score https://numer.ai/ai/toxic

I don't really see how this site is connected to real robot trading. Is it a web platform?
 
mytarmailS:

I'm in favor of it! But they're going to fuck it up anyway!

I have another question: what is the real advantage of a convolutional network over a regular one, in your opinion? in the marketplace, of course)

The convolutional networks are sharpened to classify data represented by matrices. I don't see any advantage over deep and I think LSTM is a more appropriate model since our data is a time series.

The example is offered for parsing, because there is code that can be commented, and there are some fundamental errors that will be useful for everyone to know. If of course it is of interest to the author.

Good luck

 
Vladimir Perervenko:
mytarmailS:

I'm in favor of it! But they're going to fuck it up anyway!

I have another question: what is the real advantage of a convolutional network over a regular network in your opinion? on the market, of course)

The convolutional networks are sharpened to classify data represented by matrices. I don't see any advantage over deep and I think LSTM is a more appropriate model since our data is a time series.

The example is offered for parsing, because there is code that can be commented, and there are some fundamental errors that will be useful for everyone to know. If of course it is of interest to the author.

Good luck

In fact everyone is interested, not only the author ...

I also heard that the convolutional network has a property of scalability in the recognition, that is able to recognize an object (pattern) even if it is a bit different size and shape than the training sample, is it true?

 
Dr.Trader:

And I also wonder what principle they use to create predictors. I usually have dependencies that fade with time, i.e. I trained the model on the first half of the file, and expected that on the second half of the file the error would grow over time, as usual. But there - no, everything is so stationary, without degradation over time. Although they promised trade-related data. Unusual. It is possible that the rows in the table are randomly shuffled.

Read their articles on the blog, there's no more than half an hour of pretty interesting reading, how they obfuscated the data, why all of this and so on. The strings are definitely mixed up, the attributes and data sources are classified, in addition these attributes are projected into a specific basis to mix them with each other to a state of homogeneity so you don't understand their origin.

 
mytarmailS:

In fact everyone is interested, not only the author...

I also heard the idea that the convolutional network has the property of scalability in the recognition, that is able to recognize the object (pattern) even if it will be a bit different size and shape than the training sample, is this true?


If we're talking about an image, yes, it's true.

 
Vladimir Perervenko:


If we're talking about the image, then yes, it's true.

But what about the time series?

 
mytarmailS:

What's the deal with the time line?

I didn't get good results. You need predictors (unlike other models) that are highly correlated.

I haven't found any that would give acceptable results. However I haven't experimented long. I don't have much time. Please try it. The sample has a fully workable code.

Good luck

PS. If you try it, in the input matrices, the predictors should be rows not columns.