What to feed to the input of the neural network? Your ideas... - page 19
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Firstly, there is a skewed probability of 10/20 and 99% neuronka does not give (at best from 50-70%). 99 or 100 only when averaging. And we exclude that.
I'm confused.
I would recommend you to forget about mql for now while you are in the search phase....
You need a language for rapid prototyping and with the right tools.
Then in a couple of days of learning and a couple of days of coding you will be able to run an algorithm that will create and search for ideas and then you will be able to test more ideas in one day than in a lifetime of coding in mql.
Thanks for the tips
I will leave this field to you and Maxim in the MO thread
What if you give a hammer to a robot?
There's a Japanese robot, let's say. It already plays basketball in Tokyo, communicates with little Japanese kids, has learnt to walk, answer, smarten up, etc. He sees, hears, reacts.
And then they give him a hammer. He takes it and drives nails with it.
....
And here they put him in a soft computer chair at a computer table with three monitors and say: "There, see? - There's all sorts of graphs. You've got to make money on them! Figure it out."
And so, with the participation of specialists who taught him everything, he starts to learn how to trade: poke buttons and so on. And since he is also a language model, he also learns from publicly available thematic articles, information about techanalysis, price action, etc.
In the end, everything comes down to reinforcement learning. Work on forex turns into a game, where the robot will trade for a long time on a demo in a manual tester, drawing all sorts of zones, marking peaks and so on. And, improve the skill.
This is where I am curious about DQN, because the machine will learn to trade, systematically, without memorising the price path. Like a beginner who is told in a course "close the right side of the chart and analyse on history"
I forgot to say that the probability of closing a trade in plus is not the main thing. The main thing is what SL/TP should be. And this is the main thing for neuronics. In short, there are more questions than answers. But one thing is clear SL=TP.
In my memory, all ready neuroncs as a product (market and on the side) had the same property - their SL was always larger than TP, but not much.
Even if there is no SL, an unsuccessful closing on a reverse signal was accompanied by losses, which on average exceeded 1.5 times the closing profit.
Well, well, well, how much a neuron gives to whom is up to the receiver. Stories that probability is not important, but only TP/SL ratio is important are for those children who don't know what mathematical expectation is.
It's just my opinion, but SL=TP. Spread violates this equality, and therefore, if neuronka gives 0.5+-, i.e. statistically there is no way to win, this currency pair is excluded from the portfolio.
Note that stops are present and shake the balance curve, BUT there is no averaging!
DQN is q-learning with a neural network inside instead of a table. That is, a simple algorithm. There were even transformers and more tricky models in RL articles. The result is always the same, for known reasons. For this you need to understand matstat. You will be guessing for a very long time. There are no specific ready-made tools for Forex. You will have to prepare the algorithm yourself in the end.
The implementation seemed strange. I could not read the articles at all, as if the author speaks a higher language. But, the description of the technology: "a neural network in which there are many outputs and each is responsible for a specific action."
Where are these actions? Why is there no action to buy 0.6 lots? And not 0.01. Why there is no action "Close part of position, leave 0.07 lot in the market", why there is no action "open an additional position, risk is justified".
Dynamic lot, dynamic TP-SL (or reverse signal), position maintenance - there is nothing.
Instead - buy-sell. Everything.
Although, recently there was an article "Moral Expectation at Forex". Even there they tried to explain somehow mathematically the usefulness of dynamic lot and dynamic TP-SL. After all, a trader does all this.
Therefore, I would like to see this implementation - to make sure that everything does not work to hell. I'm going to study matstat and the essence of MO.
Note that stops are present and shake the balance curve, BUT there is no averaging!
I forgot to ask you: and this is... forward, right?
Note that stops are present and shake the balance curve, BUT there is no averaging!
We all laughed, thank you.
It's not averaging, it's just masking the fit.
--
what happens to the balance when the volume drops?
and at the same increase?
to optimise the algorithm - only fixed volumes, no averaging and no locks. otherwise you will get this and that.
I forgot to ask you: is that a... a forward, right?
It's a tester. And in a tester, you can show any kind of beauty.
Unfortunately, that's true.
======
The easiest way to feed digitised bars to the neural network input. Simple and convenient.