What to feed to the input of the neural network? Your ideas... - page 37

New comment
 
Andrey Dik #:
I guess, this is the result without filters, for example, by time?

Hmm...

I, of course, have not tested with filters

But the point is that the trades are medium-term or intraday. The influence of the time filter does not seem to be strong.

But, in any case, it is possible, of course.

 
Ivan Butko #:

Hmm...

I have not tested with filters

But the point is that the trades are medium-term or intraday. The influence of the time filter does not seem to be strong.

But, in any case - you can, of course.

No, time - as an example, there are many variants of what filters to use (after training I mean).
 

An interesting phenomenon:


If you use the same architecture (any architecture), then on EURUSD, when using Softmax, the optimisation always gives top sets where BUY is not used at all. Everything sticks only to SELL and HOLD.


I look at the chart for 12 years - it is descending. That is, the MT5 optimiser adjusts the weights at the macro level to fit the long term. As a result, BUY-eval upward trend areas remain "untraded" properly.



But if you use a regular tangent as an exit, there are no problems. The top sets trade both buy and sell.

So I have abandoned this softmax, now I am digging into architectures and inputs only with tangent.



UPD

Another thing - trades up and down.




And top sets hold for a long time


 
Ivan Butko #:

Interesting phenomenon:

...


And the top sets last a long time

MLP architecture? I take it the weights are optimised by the in-house optimiser? What is the optimisation criterion, if not secret? And if the same mesh, but with some backprope algo, all other things being equal?

 
Andrey Dik #:

MLP architecture? So I understand the weights are optimised by the in-house optimiser? What is the optimisation criterion, if not secret? And if the same grid but with some backprope algorithm is used with all other things being equal?

Too many clever words

All I understand is MLP))))))

Yes, the EA is from an old article about a rudimentary neural network.

The weights are optimised by the optimiser, the input is zigzag, it turned out to be one of the most stable.

UPD

Ah, criterion - yes, the usual maximum of the complex criterion.

 
Ivan Butko #:

And the top sets last a long time

What if on top of that you train the second model again?

 
Aleksey Vyazmikin #:

What if you over train the second model again?

Please specify in the context of a regular MT5 optimiser and a regular EA.

How would it look like?

Take two sets from the optimisation list (uncorrelated), combine them and run? Or is something else in mind

 
Ivan Butko #:
Too many clever words

All I realised was that it was MLP)))))

Yeah, the EA is from an old article about a rudimentary neural network.

The weights are adjusted by the optimiser, the input is a zigzag, it turned out to be one of the most stable ones

UPD

Ah, the criterion - yes the usual maximum of the complex criterion.

Experts from the MO branch claim that:
1. You cannot search for a global extremum on a fitness function.

2. Global search optimisation algorithms (the standard GA is one of them) are not suitable for neural networks, you should use all sorts of gradient backprops for this purpose.

In short, from the looks of it, you are doing it all wrong (according to the ideas of the experts of the MO branch).

 
Andrey Dik #:

Experts from the MO branch argue that:
1. One cannot look for a global extremum on the fitness function.

2. Global search optimisation algorithms (this includes the standard GA) are not suitable for neural networks, you should use all sorts of gradient backprops for this.

In short, from the looks of it, you are doing it all wrong (according to the ideas of the experts of the MO branch).

So I will win this battle

 
Ivan Butko #:

Then I will win this battle


I think so.
1...303132333435363738394041424344...68
New comment