Machine learning in trading: theory, models, practice and algo-trading - page 2034
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
If you have experience in traditional networks, you must have tried to teach with MQL and pure C++ code just for fun, if not, try it and you will immediately see the light.
Since the influence of one ordinary trader on the market is negligible, we get what in game theory is called "playing with nature. This is all the same matstat, machine learning, and non-stationarity problems.
An approximate formal model is not difficult to build. It is possible to approximate a discrete time - no one will change position too often. We get a repeating game (that's a game-theory term) of identical rounds, with a minority winning each round-it seems like even or odd, but it's not. I then make the assumption (which I am not prepared to prove mathematically) that the resulting repeating game also has a symmetric equilibrium, which is constructed as a sequence of equilibria for games-rounds. That is, all players in each round flip coins and only win if they are outnumbered.
Greetings!
https://www.mql5.com/ru/forum/86386/page1682#comment_15839188
I can barely find this formalization of the problem, I haven't seen a better one yet.
Another question: - what is a trading strategy?
If it is easy, I would like to hear the definition of an abstract trading system
I seem to remember a promise to add WinML with ONNX)
It was written somewhere, yes...
You have one network with different types of layers, i.e. you stack them sequentially, e.g. CNN, then LSTM, then linear layer, etc. And you teach them all at once, not individually
the example I gave you - there are 2 lstm layers, then a softmax function to convert lstm outputs to the range 0;1 (otherwise it gives out -1;1), then a linear layer (to put all LSTM outputs into one), then a sigmoid at the end. I.e., now you can try CNN plugged in before the lstm still.
How to do it right - probably one needs some experience and reading some books.
https://machinelearningmastery.com/sequence-classification-lstm-recurrent-neural-networks-python-keras/Greetings!
https://www.mql5.com/ru/forum/86386/page1682#comment_15839188
I hardly found this formalization of the problem, I haven't seen any better
One more question: what is a trading strategy?
If it's not a problem, I would like to hear the definition of some abstract trading strategy
The question is fair because this approach belongs to agent-based modeling methods and it is necessary to describe the rules of individual behavior of agents. The rules of "minority game" describe only "remuneration" received by an agent from the environment.
Scientific articles on this topic usually do not ask the question "how to create a profitable system?") It rather sounds like "how exactly do these idiot agents create crises in the market?") Therefore, what they call "trading strategies" looks rather cheesy.
If to the question of what is the TS seriously and try to combine the trader's and scientific approach, the formalization of this concept escapes me. The obvious definition through the concept of algorithm suggests itself. But if you look attentively at the whole life cycle of the TS, there appear ideas like "an algorithm of reconfiguration of the TS algorithm reoptimization" and so on to infinity).
In scientific articles on this topic, as a rule, the question is not "how to create a profitable system?") Rather, it sounds like "how exactly do these idiot agents create crises in the market?") Therefore, what they call "trading strategies" looks pretty lame.
we will consider the situation with the search of information on the basis of its reliability and usefulness in the future,
The main reason for this is the lack of a clear understanding of the "scientific articles".
If we approach the question of what is a TS seriously and try to combine the trader's and scientific approach, then the formalization of this concept eludes me.
I think that's "the trick" - a clever mix of folklore (a trader's approach + grail parables) and scientific methods of time series analysis that do not take into account the nature of the market - competition + volatility spikes caused by the news background - (minus) insider information (that cannot be formalized with mathematical methods! ..... unless you remember who did their homework in math at school and who did it themselves))) )
It begs the obvious definition through the concept of algorithm. But if you look carefully at the whole life cycle of TS, you get ideas like "algorithm of reconfiguration of TS algorithm overoptimization" and so on to infinity)
OK, yes
not much, but this is the most plausible description of an abstract TS, or rather fits the problem statement of finding a TS
Then the questions again:
- how long is the life cycle of TS ? (the trader's folklore about testing it in 10 years and 10 nights, "sucked" out of a trader's finger, who turns everybody's knuckles - we should not take it into account)
- What are the optimization/reconfiguration tasks ?
Writing neural networks in the terminal is not an option at all. There any function can suddenly work in a different way than expected. Use ready and tested.
In general a different experience, everything is only in the terminal and in one code/file, the functions are simplified as much as possible. The coincidence of the calculations in the indicator and the Expert Advisor is checked. Although you are right in some respects and now you are giving me some sad thoughts ("I am right, I am wrong").
Well, I've learned from experience that bindings also glitch and slow down(
We will get through this)
In general, a different experience, everything is only in the terminal and in one code/file, the functions are simplified as much as possible. I check the coincidence of calculations in the indicator and in the Expert Advisor. Although you are right in some respects and are now giving rise to depressing thoughts (see below).
Well, I've learned from experience that bindings also glitch and slow down(
We will get through this.)
Writing a neural network from scratch just to see how it learns - doubtful fun). When you can test it on ready-made ones and not suffer from nonsense.
Also you have to write parallelization, a quality optimizer, GPU support and make it scalable. All this for the sake of understanding that NS don't work at Forex
And then you have to state that NS takes 24 hours to learn (like in recent articles) and that no research can be conducted on such architecture (due to mql or God knows what else)
Writing a neural network from scratch just to see how it learns is dubious fun). When you can test it on ready-made ones and not to suffer from nonsense
Also you have to write parallelization, a quality optimizer, GPU support and make it scalable. All this for the sake of understanding that NS don't work at Forex
And then you have to state that NS takes 24 hours to learn (like in the latest articles), and that no research can be conducted on such an architecture (due to mql features or God knows what else).
Maxim, go ahead and bring it
If my working systems don't work, I'll have to go into artificial intelligence, if you're the best there, I'm lucky