Machine learning in trading: theory, models, practice and algo-trading - page 2950

New comment
 
Aleksey Vyazmikin #:

I think it's already written that you can - what's the problem?

That the model is constantly being retrained. Are you not following the conversation at all?
 
mytarmailS #:
That the model is retrained all the time. Are you not following the conversation at all?

Make a list of models, pre-trained - and test it by switching in the code.

 
mytarmailS #:
and you can put that in a Market or a strategy tester?

Stop playing the "my favourite crutches and market/tester" card.

If you want to sell real models, make new versions of products with improved models and distribute them as regular updates.

Keep in mind that nobody wants models with daily tweaks.

 
Renat Fatkhullin #:


Keep in mind that models with a daily fit are no one's idea of a good fit.

What time interval would you recommend for retraining?
If a model has been trained in history for three years, for example.

 
Renat Fatkhullin #:

Well stop playing the "my favourite crutches and market/tester" card.

I'm just interested, nothing more

Renat Fatkhullin #:

If you want to sell real models, make new versions of products with improved models and distribute as standard updates.

I retrain 30 models every 5 minutes to make you understand and this is just the beginning....

Renat Fatkhullin #:

Keep in mind that nobody needs models with daily tweaks.

Nobody needs them? Did you make a worldwide survey that you speak for everyone? I need, for example.

Nobody needs models that don't work, that's a fact.


You are a programmer, you know how to write code, but it doesn't make you an expert in machine learning, research, working models, and it's obvious as soon as you start philosophising about what models anyone needs, or just cover one place because you understand that you can't give what they want, so it's better to say that "nobody needs it". ))

 
mytarmailS #:

You are a programmer, you can write code, but this does not make you an expert in machine learning, research, working models, and this is evident as soon as you start philosophising about what models someone needs, or just cover one place because you realise that you can't give what they want, so it's better to say that "nobody needs it". ))

I've spent enough time in this industry to understand the "value" of retraining every 5 minutes.

And I recommend getting to know a little better who I am and what I've done.

 
mytarmailS #:

And where can I read about your progress on machine learning in relation to the marketplace?

Everything created by MetaQuotes, including the company itself, is all created under my supervision and according to my ideas. Of course, with our development teams.

The value of the experience of creating and developing such an ecosystem is many orders of magnitude higher than "successes in machine learning relative to the market". And yes, it was me who brought machine learning to reality in MQL5 and MetaTrader 5. As well as Python integration.

So don't tell me about fairy tales with retraining every 5 minutes. You can fool yourself, but it's clear to the sane - it's a banal superfitting. What most simple/medium neural networks for trading systems are.


I have shown my achievements - they are visible to everyone. But yours are quite bad - you have only comments on this forum.

Do you have anything to show, any code publicly available? Do you have anything tangible?

 
mytarmailS #:
Ahhh, is that it? We delete rooms when we have nothing to say.

Show us your accomplishments, please.

You have nothing but bravado. The superfrequent overfitting is already clear.

 
Roman #:

You are told that every five minutes is a super fit.
I am also of the same opinion.

An ideal model should be trained only once!
But because we work with continuous values, the model has to be trained.
But every five minutes is really a super fit.

Even every day of retraining is not a good model.

Roman, I've been doing this for 7 years, more than 1000 scripts with different MO algorithms behind me....
I'm the one who can tell someone something here, not me...

You're saying retraining every 5 minutes is a super fit,
okay, but once a week isn't a super fit?

And what's the difference between the two?

And why is training once a week better than training every 5 minutes, and is it better?

Think about it, try to give yourself an answer, Argue yourself with this answer....


Or if you don't want to think about it at all, here's an example...
We have a moving average, is it better to recalculate the moving average once every 5 minutes or once a week?
And which values would be more relevant..

Think about it.
Why once every 5 minutes is a fit and once a week is not a fit.)))
Or both are adjustments, only in the second case with a delay of a week.


 
mytarmailS #:
Roman, I've been doing this for 7 years, I have more than 1000 scripts with different MO algorithms behind me....
I am the one who can tell someone something here, not me....

You say retraining every 5 minutes is a super fit,
Okay, but once a week isn't a super fit?

And what's the difference between the two?

And why is it better to train once a week than to train once every 5 minutes, and is it better?

Think about it, try to give yourself an answer, Argue yourself for this answer....


Or if you don't want to think about it at all, here's an example...
We have a moving average, it's better to recalculate the moving average every 5 minutes or once a week.
And which values would be more relevant..

Think about it...
Why once every 5 minutes is a fit, but once a week is not a fit)))
Or is it both, only in the second case with a delay of a week.


The difference is that the whole meaning of the concept of "machine learning" is lost.
The point is that the moving average should be trained once over a long period of time, and the model should produce accurate results.
Such a model fits into the concept of "machine learning".

And retraining every five minutes fits into the concept of "constant adjustment to the current situation".

1...294329442945294629472948294929502951295229532954295529562957...3656
New comment