Machine learning in trading: theory, models, practice and algo-trading - page 2963

New comment
 
Andrey Dik #:

what's curious? so it was said a couple of months ago in this thread in dialogues with my participation)) here many argued that max/min ff should not be in any way))))

as you set the ff so the ship will sail....

You still don't understand: it's not about optimisation at all - it's a tool to create a teacher for learning.

There are only a few problems in MOE and the first one is a good teacher.

And the quality of the teacher is NOT determined by the quality of the FF and is not determined by the quality of the optimisation - earlier we were talking about local/global optimum. In the example we didn't bother and took the first algorithm and used it head-on, which is absolutely right.

The quality of the teacher is determined by the ability to select predictors that will not lose their predictive power in the future. But when determining this property of the teacher, FF is not used at all.

 
СанСаныч Фоменко #:

You still don't understand: it's not about optimisation at all - it's a tool to create a teacher for learning. There are only a few problems in MOE, and the first one is an intelligent teacher.

No. You still don't get it. Without optimisation there can be no MoE in principle. without optimisation nothing can exist at all. nothing at all.

What is a "clever teacher"? Where are the criteria of conformity to this clever teacher? You are going to evaluate the conformity of the model to the teacher. The evaluation of conformity is the FF, how can you not understand elementary things.

The FF is the evaluation criteria. Optimisation means maximising the evaluation criteria.
 
Andrey Dik #:

No. You don't understand, you don't understand. Without optimisation, there can be no MO without optimisation. without optimisation, nothing can exist at all. nothing at all.

What is a "clever teacher"? Where are the criteria of conformity to this clever teacher? You are going to evaluate the conformity of the model to the teacher. The evaluation of conformity is the FF, how can you not understand elementary things.

The FF is the evaluation criteria. Optimisation means maximising the evaluation criteria.

While you were writing, I added to it.

I'll add.

Using optimisation when creating a teacher is one thing, is external to the model.

Using optimisation when searching for model parameters is another, it is built into the model itself, there are models where you can choose the optimisation option.

Evaluating the use of the classification model is the third and there is no smell of optimisation here. There is its own system of classification error estimation.

For example, the error matrix



More meaningful



There are special packages that extend the above examples of evaluation of classification models. For example, PerformanceAnalytics package
 
СанСаныч Фоменко #:

As you didn't understand anything, you still don't understand: it's not about optimisation itself AT ALL - it's a tool to create a teacher for learning.

There are only a few problems in MOE, and the first one is an intelligent teacher.

And the quality of the teacher is NOT determined by the quality of the FF and is not determined by the quality of the optimisation - earlier we were talking about local/global optimum. In the example, we didn't bother and took the first algorithm we could find and used it head-on, which is absolutely correct.

The quality of a teacher is determined by the ability to select predictors that will not lose their predictive power in the future. But when determining this property of a teacher, FF is not used at all.


Optimisation is used at any stage. including the stage of selecting a teacher and its parameters. then at the stage of matching the teacher - error minimisation. classification - also without optimisation.
There is not a single SIMPLE process without optimisation, and even less so in MoE.
Any optimisation is based on max/min FF.
meaningful processes are all subject to optimisation.
Non-sense processes - not all of them.
 
Andrey Dik #:

Optimisation is used at any stage. including the stage of selecting a teacher and his/her parameters. then at the stage of matching the teacher - minimising error. classification - also without optimisation.
There is not a single SIMPLE process without optimisation, and even less so in MOE.
Any optimisation is based on max/min FF.
meaningful processes are all subject to optimisation.
Non-sense processes - not all.
There's kind of a fork in the road here. The topic under discussion has turned to reinforcement learning, and maybe it makes sense to use all approaches from there. In trading it has proved itself to be about nothing so far. Either it is training with a teacher using ready-made examples. Or unknown crossed rhinos of this and that, but there is not even any theoretical apparatus for it :)

If it is pure learning with a teacher, it is more logical to look for TCs/signals and try to learn from their examples.
 
Maxim Dmitrievsky #:
It's kind of a fork in the road here. The topic under discussion has turned to reinforcement learning, and maybe it makes sense to use all approaches from there. In trading, it has proved itself to be about nothing so far. Either it is training with a teacher using ready-made examples. Or unknown crossed rhinos of this and that, but there is not even any theoretical apparatus for it :)

If it is pure learning with a teacher, it is more logical to look for TCs/signals and try to learn from their examples.

You can do it that way. ten years ago I showed you how to find ideal inputs for a TS on BP - what is not a teacher? - The sensitivity threshold in this case is spread and commission, the trading frequency will depend on the threshold.
I have not done any research, but it would be interesting to see the statistical characteristics of such an "optimal series" as a teacher.
 
Andrey Dik #:

You can do it that way. ten years ago I showed you how to find ideal entries for a TS on BP - what is not a teacher? - The sensitivity threshold in this case is spread and commission, the trading frequency will depend on the threshold.

Here is https://www.mql5.com/ru/code/903 also 11 years ago. And the chart is prettier than mytarmailS in his example.

But I don't think it's yours. Where is yours?

Sampler
Sampler
  • www.mql5.com
Индикатор i_Sampler рассчитывает идеальные входы, предназначен для обучения нейросети.
 
Forester #:

Here's https://www.mql5.com/ru/code/903 also from 11 years ago. And the graph is prettier than mytarmailS in his example.

But it doesn't look like yours. Where is yours?

 
Forester #:

Here's https://www.mql5.com/ru/code/903 also from 11 years ago. And the graph is prettier than mytarmailS in his example.

Here it is before training, and his is after. That is, it not just plots nicely, but also brings the model to a minimum error. The approach is interesting, but it has not been worked out how to improve it on new data.
 
Maxim Dmitrievsky #:
Here before the training, and he has it after. So he doesn't just mark it up nicely
Exactly.

Not only that, I have a balance chart for 100 candles 5min, that's less than a day, and this one has a balance for half a year

The question is: where would the perfect balance look smoother?

Really, I'm writing this and I'm embarrassed to have to explain it.

1...295629572958295929602961296229632964296529662967296829692970...3700
New comment