Machine learning in trading: theory, models, practice and algo-trading - page 3731

New comment
[Deleted]  
Forester #:
What about the signal? It's beautiful in the tester. I wonder how it will be in real life.
Who should I make the signal for, for myself? I'll put it on my account, of course. It'll be worse in real life.
[Deleted]  

Distribution of errors by hours for the channel ML strategy on the minutes, where the cluster corresponds to a specific trading hour:

Iteration: 0, Cluster: 0
R2: 0.9958055181204825
Iteration: 0, Cluster: 1
R2: -0.9198035778567698
Iteration: 0, Cluster: 2
R2: -0.9405239683877826
Iteration: 0, Cluster: 3
R2: -0.9247647899181217
Iteration: 0, Cluster: 4
R2: -0.9824998494931588
Iteration: 0, Cluster: 5
R2: -0.9895731405068378
Iteration: 0, Cluster: 6
R2: -0.9905500842939968
Iteration: 0, Cluster: 7
R2: -0.9851648150363834
Iteration: 0, Cluster: 8
R2: -0.9042140030854123
Iteration: 0, Cluster: 9
R2: -0.8922150254448274
Iteration: 0, Cluster: 10
R2: -0.9435129811726439
Iteration: 0, Cluster: 11
R2: -0.9602763137808074
Iteration: 0, Cluster: 12
R2: -0.9925125476619788
Iteration: 0, Cluster: 13
R2: -0.8634130934563771
Iteration: 0, Cluster: 14
R2: -0.8513933806661487
Iteration: 0, Cluster: 15
R2: 0.6585349517759875
Iteration: 0, Cluster: 16
R2: 0.514241043820456
Iteration: 0, Cluster: 17
R2: -0.02676637285513661
Iteration: 0, Cluster: 18
R2: 0.3641192892793419
Iteration: 0, Cluster: 19
R2: 0.9696362778162367
Iteration: 0, Cluster: 20
R2: 0.9539071315308307
Iteration: 0, Cluster: 21
R2: 0.9501832578880937
Iteration: 0, Cluster: 22
R2: 0.9920145219103729
Iteration: 0, Cluster: 23
R2: 0.9922260913155719

Made possible by a fast tester

 
Maxim Dmitrievsky #:

Distribution of errors by hour for the channel ML strategy on the minutes, where the cluster corresponds to a specific trading hour:

Throwing this clock out of the trade?


This failure is observed on almost all symbols. It was a tweak to Trump. From my point of view, this piece should be thrown out completely when training. Right now it is in OOS, but next year it will definitely fall into the learning interval, and then you have to throw it out. But even when assessing robustness at OOS, it is worth ignoring the performance at this point (the Spring 2025 slice).


I assume that there have been such intervals before.

[Deleted]  
fxsaber #:

Kicked that watch out of the trade?


This failure is seen on virtually all characters. It was a tweak to Trump. From my point of view, in training this piece should be completely thrown out. It's in OOS now, but next year it will definitely fall into the training interval, and then we should throw it out. But even when assessing robustness on OOS, it's worth ignoring the performance at this point (the Spring 2025 piece).


I assume that there have been such intervals before.

Yes, the reds are getting thrown out. I'm still running it back and forth in the tester, I think I can still improve it.

There from training to training you get this dip, sometimes there is no dip. Poorly controlled process :)

 
Maxim Dmitrievsky #:

There from training to training you get this failure, sometimes there is none.

When I encounter something like that there, I take it as luck.
 
Maxim Dmitrievsky #:

Made possible by the quick tester

There are algorithms that allow you to calculate throw away intervals almost for free. And not roughly by hours, but precisely.
[Deleted]  
fxsaber #:
There are algorithms that allow you to calculate throw away intervals almost for free. And not roughly by hours, but precisely.
I've heard about such algorithms, we can think about how to integrate them into our code so that it would be beautiful. Because the whole basis is in Python. There are filters not necessarily by time, volatility filters work too.
 
Maxim Dmitrievsky #:
I've heard about them, we can think how to integrate them into our code to make it beautiful. Because the whole basis is in Python. There are filters not necessarily by time, volatility filters work too.
The algorithm of throwing out, of course, is the same for any input data.
[Deleted]  
fxsaber #:
The algorithm of discarding is of course the same for any input data.
Yes, but the tasks are a bit different in the case of post facto discarding from a ready TC and discarding initially without a TC
 
Maxim Dmitrievsky #:
Yes, but the tasks are slightly different in the case of throwing out ex post facto from a finished tc and throwing out initially without a tc

Almost the same tasks

  • Optim TS, where among customisable weights there is also a filter.
  • Optim TC without filters, where the filter is selected at the end of each pass.
1...372437253726372737283729373037313732373337343735373637373738...3741
New comment