Machine learning in trading: theory, models, practice and algo-trading - page 2643

 
Maxim Dmitrievsky #:
It's about the same, a little more.
It's probably a question of overfit, the new data will show different curves.
The third piece of the picture is the new data.
 
Maxim Dmitrievsky #:
Interesting, I'll try to think about it later, it's a bloody mary today, it's hard to think.
Or train model1 so that it outputs a useful feature for model2.

There's no limit to my imagination.
 
Maxim Dmitrievsky #:
Remember a long time ago when we were successfully forecasting clusters. Do you remember how we did targeting?
 
mytarmailS #:
Remember a long time ago when we successfully predicted clusters. Do you remember how we did targeting?
I predicted the cluster number through a tree or a multiclass forest. I also used to mark up the targets with clustering.
 

Well, yes, I am interested in marking targets, but not to predict the cluster on the next candle, but let's say we have the current hour (5m tf) and forecast what cluster will be the next hour, did you do that? Do you remember?

 
mytarmailS #:

Well, yes, I am interested in marking targets, but not to predict the cluster on the next candle, but let's say we have the current hour (5m tf) and forecast what cluster will be for the next hour, did you do that? Do you remember?

I think I did it on the next candle, but if you get clusters of many candles in a row, you can do it for the distant future.

The source code was lost when I moved to a new computer, the old one had a glitch, I didn't save it in the cloud.

I remember that if you mark up through clustering of increments, the results are very stable on new data, but the marks themselves are so bad

Ah, I had something in Google colab, I can pull it up if you need it.
 
Maxim Dmitrievsky #

Oh, I've got some left over from Google colab, I can pull it up if you need it.
No, I'm good. I'll write it myself.
 
Aleksey Nikolayev #:

If you want to add continuous time, it is already a generalisation of Markov chains -- semi-Markov models (processes).

I am not ready to promise help, but I can participate in open discussion of the topic as far as possible.

In what sense "continuous time" - the point is that we have events (time scale) in the form of a signal to enter the market, and there are "patterns" that were present or not at the moment of signal appearance. Therefore, there will be moments when there is a point on the timeline, but there is no pattern. I think that the time intervals of occurrence (absence of n discrete periods) of the pattern are also important to take into account its influence.

There is also such a thing, because the number of patterns can be more on the whole history interval than it was at the moment of the signal, then maybe it should be taken into account, i.e. how many per cent of patterns accompany the signal, because if it is a small percentage, then the connection with the signal is either random, or the signal is too filtered by the basic condition. However, there is a discreteness issue here - a pattern can exist continuously for n bars in a row. I think that there should be discreteness, maybe for the same ZZ, then if the signal and the pattern are the same, then additional statistics is meaningless, and if not, it can be useful.

Thanks for your willingness to help, albeit in a limited amount! I haven't started the code yet - I want to finish another project, but it is useful to think about this topic for future experiments.

 

Let's get some research published, some ideas.

 
Here, even if you publish a ready-made grail, in response they will start explaining to the author what a fool he is)