Machine learning in trading: theory, models, practice and algo-trading - page 341

 
SanSanych Fomenko:


It's not about R


I have copied your article on Rattle, I am not quite sure why you also fed the bare prices there, but never mind, it was a good article) I ran your example and got exactly the same results. I still have a question about RNN - advise me some good package, if available in R, in particular, LSTM. Vladimir already wrote me that it's better to use Python for complicated NS, but R has it ) It turns out, it's quite easy to use

p.s. And once again, please send me info about your training course

 
Maxim Dmitrievsky:


I've read your article on Rattle, I don't really understand why you added the bare prices there too, but never mind, it's a good article) I ran your example and got exactly the same results. I still have a question about RNN - advise me some good package, if available in R, in particular, LSTM. Vladimir already wrote me that it's better to use Python for complicated NS, but R has it ) It turns out, it is quite easy to use

p.s. And once again, please send me info about your classes


I can not help: I am not involved in networks, my experience is only with nnet in rattle. This experience is negative.

I don't have a training course.

My article intentionally contains a rather large set of predictors. If you learn how to cull them, you can get the error OUT OF THE TRAINING FILE (this is not an OOV) on the article predictors down to less than 35% for scaffolding and ada.

Good luck

 
Maxim Dmitrievsky:


I have copied your article on Rattle, not sure why you put the bare prices there, but never mind, it's a good article) I ran your example and got exactly the same results. I still have a question about RNN - advise me some good package, if available in R, in particular, LSTM. Vladimir already wrote me that it's better to use Python for complicated NS, but R has it ) It turns out, it is quite easy to use

p.s. And once again, please send me info about your classes

If without Python - rnn and mxnet . Purely in R.

Good luck

 
SanSanych Fomenko:


Can't help: I don't do networks - only experience with nnet in rattle. That experience is negative.

I don't have a training course.

My article intentionally contains a rather large set of predictors. If you learn how to cull them, you can get the error OUT OF THE TRAINING FILE (this is not an OOV) on the article predictors down to less than 35% for scaffolding and ada.

Good luck


Vladimir Perervenko:

If without Pythona, rnn and mxnet. Purely in R.

Good luck


thank you :)
 
Vladimir Perervenko:

You're building your sentence wrong. You write: "I could not find the filters I needed. Since I don't know which filters you are interested in, here are a few at a glance:

mFilter package - Baxter-King filter, Butterworth filter, Christiano-Fitzgerald filter, Hodrick-Prescott filter, Trigonometric regression filter

FKF package - Fast Kalman filter ....

Also, if you are deep into the subject of filters and know the mathematical formula by which it is calculated, no problem to just calculate it. No?

Good luck

Thanks, didn't know that.

If you know the math formula... Who knows the formula?) Filters are designed for a specific task, and it is not enough to take Bessel or Kalman patterns and apply them. You also need tools for working with filters. Filters are not always used in their original form.

I had thoughts to use R and SciLab together, but marshalling R<->SciLab data is a rather laborious task and unlikely to make sense, at least in the development stage.

 
Yuriy Asaulenko:

Thank you, I didn't know.

If you know the math formula... Who knows the formula?) Filters are designed for a specific task, and it is not enough to take Bessel or Kalman patterns and apply them. You also need tools to work with filters.

I had a thought of using R and SciLab together, but marshalling data R<->SciLab is rather labor-intensive task, and it hardly makes sense, at least in the development stage.

Feel free to contact me.

If you use Matlab, marshalling R<-> Matlab is a done deal.

Good luck

 
SanSanych Fomenko:

But in reality the pseudo-problem of R-filters you identified has much deeper roots.

Why do you need them? The filter is an auxiliary tool. And R has ready-made solutions for building decision-making blocks. We can designate two main lines: machine learning and ARMA-ARIMA-ARFIMA-ARCH-GARCH. And what does this have to do with filters per se?

Why filters?

There is a field called statistical radio engineering. Simply put, it is the science of detecting and isolating signals from noise, and even from under noise.

In the general case, before detecting or recognizing a signal, we must by all kinds of transformations strengthen the energy spectrum of the signal and weaken the noise components, which will naturally increase the signal to noise ratio of the time series and simplify further signal processing and recognition.

In our case, what is a signal and what is noise is up to everyone, depending on the strategy.

 
Yuriy Asaulenko:

Why filters?

There is a field called statistical radio engineering. In simpler terms, it is the science of detecting and isolating signals from noise, and even from under noise.

In the general case, before detecting or recognizing a signal, we must by all kinds of transformations strengthen the energy spectrum of the signal and weaken the noise components, which will naturally increase the signal to noise ratio of the time series and simplify further processing and recognition of the signal.

In our case, what is a signal and what is noise - it is up to everyone, depending on the strategy.


A typical mistake of radio engineers is that they think that there is a signal on financial markets and they cannot imagine even in their wildest dreams that there is no signal on financial markets and there never will be. This is why filters are almost never used in financial markets.

There is another, purely technical circumstance: financial markets are non-stationary time series, as a result a lion's share of statistics that work fine in radio engineering goes down the drain. I have outlived so many radio engineers on this forum. I advised them all: if you want to make money, forget radio engineering. For ever.

 
SanSanych Fomenko:


There is no signal in the financial markets and there never will be.

I think* there is one after all. Let's say I'm trying to build a strategy by trading opening prices on H1. There are no stops or takeoffs, but only function CopyOpen() from mql, and advisor that once an hour decides where the price will be in an hour and takes a position in that direction. It turns out that I work with a signal with a sampling rate of 1/3600 Hz, no?


* I am not a radio technician and do not know the terms, perhaps the opening price should not be called a signal but something else.

 
SanSanych Fomenko:


Typical mistake of all radio engineers - they believe that there is a signal on financial markets, but even in the worst dream they cannot imagine the situation that there is no signal on financial markets and never will be. This is why filters are almost never used in financial markets.

There is another, purely technical circumstance: financial markets are non-stationary time series, and as a result, the lion's share of statistics that work well in radio engineering goes into tatters. I have outlived so many radio engineers on this forum. I advised them all: if you want to make money, forget radio engineering. Forever.

Well, if there is no signal, what are you looking for? You are looking for a signal without admitting it.)

So, what is meant by a signal in the market - a certain set of patterns (in the sense of images in some space), indicating the possibility to enter the trade at a given moment of time. A typical classification problem. And, by the way, before carrying out the classification, it is desirable to make this space if not orthogonal, then at least linearly independent, which is impossible in principle without using any filtration.

As I understand it, your predictors are an attempt to increase the signal-to-noise ratio.