What to feed to the input of the neural network? Your ideas... - page 79

New comment
 

Python models work wonders

For the first time the result appeared with a lot of inputs: I feed as many as 160 pieces. (Why 160? I just stopped at this number, for no reason)

Usually the more inputs the worse it is. But this one went well.

result of the trade as a target.


2) Train using any method.

3) Launch our (pomoichny) strategy

4) Add a filter to it - our NS.


As a result, we need to put some intellectual, creative effort, patience, time to pick a strategy.

The neural network will do the rest.

We remove from the neural network an unbearable intellectual labour simply because in the current realities the NS is not even able to determine the levels, if they are not marked "by itself".

And it will do what it is supposed to do: to grind the TS.



And by increasing the regression threshold on the forward, the inputs are also improved. Usually, in 99% of cases, increasing the input threshold does not give anything and stupidly reduces good inputs together with chaos proportionally, but here quality inputs remain, and noise goes away.

BiLSTM model (LSTM gives the same thing, the difference is small).

 
Ivan Butko #:
And here it is just prices + method of determining inputs and target by TS:
.

So just prices without any conversion? Curious!

Ivan Butko #:
Python models work wonders

This is great! There should be at least sometimes a positive from hard work!!!

 
Aleksey Vyazmikin #:

I.e. just prices without any conversion? Curious!

That's great! There must be at least sometimes a positive from hard work!!!

By habit I do not add: normalisation as usual: the whole set (160 prices) is brought into the range -1...1
 
Ivan Butko #:
I don't add it by habit: normalisation as usual: the whole set (160 prices) is brought into the range -1...1

What is the training period? Were the closing prices given? From what TF? Test on demo from MQ? What is the mat. expectation per 1 lot?

 
If it weren't for the six-month drawdowns, it would be fine. I have had them for up to 2 years. When testing with 2015 valkingforward.

PS. How can you all work at night? Daytime is the most efficient time.
 
Well at least the movement towards a healthy MO has started, not a bad thing already
 
Aleksey Vyazmikin #:

What was the period of study? Were closing prices given? From what TF? Demo test from MQ? What is the mat. expectation for 1 lot?

Switched to 21-year: 2000 to 2021.

The two plots are too irreconcilable: 2000 to 2012 and 2012 to 2021.
If they both go up: either retraining or something working.

Test on icMarkets, real account. I don't know the details, I'm not at my computer right now.

As for the selection of entry prices: I do not spread here yet, I have been digging for two years, I save the best ones. I can post something curious.
 
Forester #:
If it weren't for the six-month drawdowns, it would be fine. I have had them for up to 2 years. When testing with 2015 valkingforward.

PS. How can you all work at night? Daytime is the most efficient time.
Yes, head-on results are nil.

But I liked the fact that the "relay" works: if you turn it one way, there is a lot of noise (rubbish) on the forward, if you turn it the other way, there is less noise, but good deals are not cut off.
 
Ivan Butko #:


What's this programme?

 
Andrey Dik #:

What's this programme?

I was trying to get chat to run a python NS on my RTX 3080. He offered me a development environment: Jupiter, GoogleCollab, Idle, or Spider.

I like the look of Spider, I run the model scripts in it
1...72737475767778798081828384
New comment