Machine learning in trading: theory, models, practice and algo-trading - page 2578

 
elibrarius #:

That's a lot of deletions. What server?

I usually do all the checks on a working server and a working (not demo) account.

For EURUSD I very rarely miss bars on M1 - not every day and not more than 2-5, usually at night. And they are skipped because there were no ticks during the bar.

Other currencies with less active trade have more skips. The pound seems to be active, it is strange that you deleted 20% of bars.

Six months... what about 2-3 years?

You do the sampling, I'll check it out.

Welltrade server, like even real.
 
mytarmailS #:
Do a sample, check

Welltrade server, like even real

How do I make it? As a CSV? DucasCopy will give you at least 20 years of quotes. I guess you just need to increase the length of your sample. Yes ... and soon you will get the limit.

The MT5 tester only gives data for the current year and the last year (from the start date of testing). I've already asked them a couple of times to remove that limit, but no one is particularly supportive.

If you need more data than 1-2 years - ask too. If a lot of people ask, maybe they will.

 
mytarmailS #:

What is a perfect spread between pairs? How do I express it in numbers? These are the main questions

open question

 
mytarmailS #:

the question is open.

Co-integrated.
Cointegrated test.

 
Roman #:

Co-integrated.
Cointegrated test.

Yes, I got it, I got it))
I didn't mean it that way, I didn't phrase my question well, it's not convenient to rewrite my question on the phone now
 
Of course, the series need not be cointegrated in the classical sense of error correction, i.e., with stationary residuals. The cointegration test will only show this
 
Fast235 #:

I remember when you asked to spin the spinner)

Remind me how you were a while back in the picture and seychas?
Remind me 🤣
 
One can approximate any such series to a stationary one by resampling, through generative models, and train on it, see what happens. This is to avoid fitting a specific model, like a sliding regression. And for more uniform sampling, no outliers. Well, and do it on the increments. For a nonlinear machine learning model, that would be fine. Also work on the timings.

The question of the third "do not trade" class, which should include weakly predictable cases, has already been raised in passing here. I am sure there are beautiful ways to filter them, at the level of metrics of trained models themselves.
 
Because of the discreteness of the volume, the smallness of the deposit (and if you want to have a portfolio of spreads) will not be such a large set for possible spreads. It is possible to check them with a simple search in the tester.
 
Aleksey Nikolayev #:
Because of the discreteness of the volume, the smallness of the deposit (and if you want to have a portfolio of spreads) will not be such a large set for possible spreads. It is possible to check them with a simple search in the tester.
Indices, futures+spot mostly trade this way, with low returns, on the exchange. I don't see any successful attempts on forex. I tried something like SnP vs Dax, I do not remember exactly. It works, but no golden mountains.