What to feed to the input of the neural network? Your ideas... - page 5

New comment
 

You're just pissed off that you can't..... that's the whole ideology.

What's the big deal? People will experiment. If it's good, good. If it's not good, fuck it.

 
Dmitry Fedoseev #:

This outboard is nothing like the neural net. Nothing to do with it at all.

If we proceed from the postulate that nothing on the market does not repeat itself for a long time, then it is not worth applying any method at all.

You are thinking in the right direction! There is only one method.

It's like in the market, whoever has more money (or something else) is right.

 
Dmitry Fedoseev #:

You're just pissed off that you can't..... that's the whole ideology.

What's the big deal? People will experiment. If it's good, good. If it's not good, fuck it.

Well, let them experiment, I've experimented enough.


 
JRandomTrader #:

An outside perspective: to detect seasonality, it's a good idea to add the day of the year as well.

JRandomTrader #:

External view: it would be nice to add a day of the year to detect seasonality.

Ideally, yes. But then you need to have training for several years - averaging the same day in several periods. Otherwise it makes no sense. And memory in mt is enough for 7 months of training at best. Although, if you take a large timeframe - it is realistic. But large timeframes have long drawdowns. I hope someone will take up to test on periods from an hour.

 
Lilita Bogachkova #:

Not necessarily, the signal of the indicator may well be ahead of the events. CCI is one of such indicators. As you can see in the figure, CCI sometimes outperforms the RSI signal (signal red histogram - sell, blue histogram - buy). And if the signals of the indicators coincide, the signal is most likely correct.



There is one more disadvantage in indices - herdiness. And some people make sequences to trigger all indices in a predictable direction....

 
Dmytryi Voitukhov #:

There's another flaw in turkeys - herding. And someone puts sequences to trigger all the turkeys in a predictable direction....

there are two other drawbacks to the turkeys:

- One is that they're already part of the recognition system. They are used only in a complex to recognise this or that movement of the initial price. It is possible to teach NN to trade according to alligator principles, it may be scalable and adaptive. But this is not a popular method here. They want everything at once or just count sine through NN

- turkey is a bird :-)

 
Roffild #:
I gave a clear answer

Thank you
 
Maxim Kuznetsov #:

there are two other drawbacks to turkeys:

- The first is that they are already part of the recognition system. They are used only in a complex to recognise this or that movement of the initial price. It is possible to teach NN to trade according to the alligator principles, it may be scalable and adaptive. But this is not a popular method here. They want everything at once or just count sine through NN

- turkey is a bird :-)

In case of serious interest, I have no problem instead of prices I can drive any bird :) and give to play. And suddenly ...

 
After experimenting to increase pressure from late nights and lack of sleep, I came to one direction:

A neural network of the 5-3-5 kind on a year-long training period and on a forward equal to up to a year balances between noise and working patterns. Любая другая сеть, даже ацкий Скайнет вида 100-100-100-100-100-100-100-100-100-100 (плакал тогда мой 10700к), недообучает, либо переобучает сеть.
Undertraining occurs even with a large number of neurons and layers, otherwise it is overtraining.

And only on 5-3-5 was it possible to achieve 7 closed in + months out of 8, this year. Training was done every month, considering the new one and cutting off the old one (the first one).
As soon as it started going down into 2021, the net stopped working, namely learning. It flattens on history as well as the forward. Apparently, 2022 is full of repeating patterns.
I've been inputting... Oh, God, I forgot. The exports High[1]-Low[1] and further by chronology for 5 candlesticks in a row are saved on my computer. It must be this scheme.
I tried 24 candlesticks, 48, 72 and 486 candlesticks to input. The results are flat, 50/50, either at 0, or plum, or profit. Not good for practice.

I'm sick of neural networks, so I'll pause. Thank you all for throwing ideas and discussing some things.
I'm sure I will come back to this topic, as I did three times before.
 
Ivan Butko going to pause. Thank you all for throwing ideas and discussing things.
I'm sure I'll come back to this topic as I've come back three times before.

Then what's the point of me preparing a detailed response?

123456789101112...86
New comment