Machine learning in trading: theory, models, practice and algo-trading - page 3134

 
СанСаныч Фоменко #:
    Valeriy Yastremskiy #:
    Maxim Dmitrievsky #:

    I want to discuss about an algorithm that will search for regularities itself, so that the regularities are not in a ready-made form.

    I want it to be able to look for patterns of any kind, without bias to the data and ways of processing them.


    But I don't know what kind of (what language) algorithm can describe these regularities, and in general, the structure of regularities itself.


    Therefore, the question is whether there is a universal language that can describe anything and how to put this "anything" into the "checker" in the form of ready modules


    I'm inclined to log. rules and code in general.

     
    mytarmailS #:

    I want to speculate about an algorithm that will look for patterns on its own, so the pattern is not there in a ready-made form.

    I want it to be able to look for patterns of any kind, without bias to the data and the way it is processed.


    But I don't know what kind (what language) the algorithm can describe these regularities, and in general, the structure of regularities....


    Therefore, the question is whether there is a universal language that can describe anything and how to put this "anything" into the "checker" in the form of ready-made modules


    I'm inclined to log. rules and code in general

    The whole history of mankind says that only man can find a pattern.

    Today, greyhound writers attribute this ability to AI, which today works ONLY on extrapolation of historical data. That's at best. And usually AI is nothing more than a very fast search engine.

    You are posing a qualitatively different problem: to create an AI analogue to a human one. If you realise that, then so be it.

     
    mytarmailS #:

    I want to speculate about an algorithm that will look for patterns on its own, so the pattern is not there in a ready-made form.

    I want it to be able to look for patterns of any kind, without bias to the data and the way it is processed.

    But I don't know what kind (what language) the algorithm can describe these regularities, and in general, the structure of regularities....

    Therefore, the question is whether there is a universal language that can describe anything and how to put this "anything" into the "checker" in the form of ready modules

    I'm inclined to log. rules and code in general.

    You are late, I have already created such a module )) but only for forex or any time series.

    Data representation for different areas is different and the representation of patterns is different too, it will be a mess.

     
    mytarmailS #:

    I want to speculate about an algorithm that will look for patterns on its own, so the pattern is not there in a ready-made form.

    I want it to be able to look for patterns of any kind, without bias to the data and the way it is processed.


    But I don't know what kind (what language) the algorithm can describe these regularities, and in general, the structure of regularities....


    Therefore, the question is whether there is a universal language that can describe anything and how to put this "anything" into the "checker" in the form of ready-made modules


    I'm inclined to log. rules and code in general

    The task is not clearly given. Searching in a one-dimensional series, array is one thing, in a 2-dimensional or more is another. A one-dimensional price series is too simplistic, without taking into account spread and volume. Besides, calendar time cycles this series, which simplifies on the one hand, but it is another condition / factor that must be taken into account.

    And what kind of repeatability to look for, if we look at what we are interested in, it is trends or patterns, but there are more complex dependencies that can also repeat, but they will be useless for trading.

     

    Log rules are associative by nature. Of course, you can describe them, but you need some validator to determine the reliability of these rules.

    This is the topic of transition from association to causality. I don't know such a matrix for log rules, though probably something similar to what is done through regression.

     
    Maxim Dmitrievsky #:

    You're too late, I've already created one ))

    you have created the usual "leaderboard" models, it's like comparing a paper aeroplane with a real one.

     
    Valeriy Yastremskiy #:

    And what kind of repetitions to look for, if we look at what we are interested in, it is trends or patterns, but there are more complex dependencies, which can also be repeated, but they will be useless for trading.

    All we need to know are two classes = reversal/non-reversal.

    Everything else - patterns, trends, flatlines, news, levels - is firstly subjective and secondly it can be classified as signs.

     
    mytarmailS #:

    You've created the usual "leaderboard" models, it's like comparing a paper aeroplane to a real one.

    Leaderboard is at the end, because in real life you will never get one one and only and perfect. There will always be a difference in judgement.

     
    Maxim Dmitrievsky #:

    Leaderboard is at the end of it, because in real life, you're never going to get one one and only and perfect

    We're talking about different things, I'm basically talking about the idea of an automaton creating very complex qualitative features for a model that are not constrained by anything.


    It can be a complex rule ==> on which some sequence is trained, ==> only working patterns (rules) are selected from it ==> on which the model is trained.


    and it can be only one feature out of hundreds for the basic model.

     
    mytarmailS #:

    We're talking about different things, I'm basically talking about the idea of an automaton creating very complex qualitative features for a model that are not limited by anything

    It can be a complex rule ==> on which some sequence is trained, ==> only working patterns (rules) are selected from it ==> on which the model is trained.

    and it can be only one feature out of hundreds for the basic model.

    don't forget to get a supercomputer )