Machine learning in trading: theory, models, practice and algo-trading - page 2480

[Mihail Marchukajtes]

Dmytryi Nazarchuk #:

When applying machine learning methods to RUNNING RATES, it is almost impossible for the same set of input variables to have the same dependent variable. Different values of the dependent variable form a prediction error that must be minimized.

This entire thread is about minimizing prediction error, Axakal.

Plain truths....

Do you know that there are such methods of training where error minimization can be infinitely long not in time but in actual result and that the minimal error obtained during training is not a criterion for evaluating generalization ability of a model. Say the method of back propagation of error is a way to minimize error on any data to zero, but as a rule such models have no generalizing ability and work poorly on new data. So this method will be able to minimize the error in the presence of contradictory data to zero, but it will be of little practical use. So believe me minimization of error in training is not sufficient approach in training!

[Mihail Marchukajtes]

mytarmailS #:

why random?

Exactly. The time series of the exchange is not stationary first of all, but it is definitely not random. There is always a reason for the change of price series and it is the analysis of the reason which helps to predict the price, not the consequence!

[Dmytryi Nazarchuk]

Mihail Marchukajtes #:
Exactly. The time series of the exchange is not stationary first of all, but it is by no means random. There is always a reason for the change of the price series and it is the analysis of the reason that helps to predict the price, not the consequence!

........................ Give me an example of a deterministic nonstationary series

[Dmytryi Nazarchuk]

Mihail Marchukajtes #:
Are you aware that there are such methods of training where error minimization can be infinitely long not in time but in actual result and that minimal error obtained during training is not a criterion for evaluation of generalizability of model. Say the method of back propagation of error is a way to minimize error on any data to zero, but as a rule such models have no generalizing ability and work poorly on new data. So this method will be able to minimize the error in the presence of contradictory data to zero, but it will be of little practical use. So believe me minimizing error in training is not a sufficient approach in training!

Bullshit

[Dmytryi Nazarchuk]

mytarmailS #:

why random?

Time series are deterministic, random and stochastic. There are no others. None at all.

Forex and stock quotes - what kind of series?

[JeeyCi]

Evgeniy Ilin #:

And if it's your thought exactly, I understand what you're thinking, you can take any function like:

A[1]*X^0+A[2]*X^1+ ... + A [N]*X^N, it is in general a Taylor series (functional series), except that A[i] > 0 for all i = 1...N it gives in general a constant increase in the first derivative, to put it clearly, so

How to Difference a Time Series Dataset with Python

How To Backtest Machine Learning Models for Time Series Forecasting

Yes, there's something on the web ... I'm confused by seasonal adjustment and other more significant events in time...

... and The number of times :

As such, the process of differencing can be repeated more than once until all temporal dependence has been removed.

The number of times that differencing is performed is called the difference order.

p.s.

I will look through the links from here too (thanks for the article)

[mytarmailS]

Dmytryi Nazarchuk #:

Time series are deterministic, random and stochastic. There are no others. None at all.

Forex and stock quotes - what kind of series?

I'm not good at it.

If I'm not mistaken: from the point of view of probability theory, quotes are a random, non-stochastic process.

But I don't agree with it.

[Dmytryi Nazarchuk]

mytarmailS #:

I'm not good at this...

If I'm not mistaken: in terms of probability theory, quotes are a random, non-stationary process.

But I don't agree with that.

Why not?

[mytarmailS]

Dmytryi Nazarchuk #:

Why?

I have several arguments, but they should not be considered in terms of probability theory, but just human reasoning.

1) All mathematical methods invented for processing random/non-stationary/stationary..... any series do not work for quotations, why?

2) the process is organized by people to take money from other people, it cannot work randomly, I believe this process is deterministic but complicated...

There were other points, but when I started writing, it slipped my mind ...

[Dmytryi Nazarchuk]

mytarmailS #:

I have several arguments, but they should not be considered in terms of probability theory, but just human reasoning.

1) All mathematical methods invented for processing random/non-stationary/stationary..... any series do not work for quotations, why?

2) the process is organized by people to take money from other people, it cannot work in a random way, I believe this process is deterministic but complicated...

There were more points, but when I started writing, they slipped my mind...

1. All mathematical methods for non-stationary processes are shamanism. All mathematical methods for non-stationary processes are shamans. That's because you can predict the future only by looking at the past; if the future does not depend on the past, forecasts based on the past do not work.

So the choice of method, model, etc. does not play any role - only the right choice of input variables.

You can go no further