Machine learning in trading: theory, models, practice and algo-trading - page 1215

[Maxim Dmitrievsky]

Igor Makanu:

Yes, I told you, you were the only one who could simplify the MO task, everything is searchable, everything works, there were examples from Reshetov in the early days of MQL development, but they are primitive, though ))))

the only thing left is to lick the algorithm to the end and finish this whole thing with RL :) maybe in Python you'll get some % model quality due to higher grading quality

but it's a lot of work to do.

[Maxim Dmitrievsky]

Martin Cheguevara:

My equity chart is not random and is quite informative (I need to check it), I have learned to distinguish trends from flattens.

The trade is going on, but it needs to become more efficient.

Where is the equity graph?

[mytarmailS]

Martin Cheguevara:
I'll add page analysis via google yandex search engines to it

So why do you need it?

[Igor Makanu]

Maxim Dmitrievsky:

the only thing left is to lick the algorithm to the end and finish this theme with RL :) there may be on python due to the higher quality of classification will still be possible to squeeze out some % of model quality

I have a lot of work to do.

I tried it in VS2017 for a couple of times yesterday. Python works there and even in a separate widows-form. The only thing there is IronPython 2.7, I need to analyze it, but I think it may be compatible with MT5 without any problems

[Ivan Negreshniy]

I have learned to distinguish flat trends, but I need to improve my trading efficiency:

Yes, I just thought why create something myself, I am interested in the cause-effect relationship of the two variables my program is already able using Apache Lucene, JSOUP, JSON, Apache POI and so on technologies to recognize text anywhere in anything in the pictures to documents and so on (this is accompanied by information matrices (stored in a distributed database) according to which is indexed information recognized in graphic objects) if something can not - looking for a site to convert data into an acceptable format for recognition or itself if can transfer.

The thing is that I do not want to reinvent the wheel ... I just need to find a neural network capable of fast learning with two input variables - the equity data, and the trend indicator.

(experience in Java EE for about 5 years, there are many projects already implemented).

I'm not even trying to attach neuronics to market trading. It is unnecessary and most likely impossible at the moment, since there was not at least one implementation of a stable earning neural network.

My Equity graph is not random and quite informative (i need to check it), i have learned to distinguish trends from flattens.

I have learned how to distinguish flat trends. The trading is going on, but i need to improve performance.

I was once engaged in the creation of neurofilters to improve the efficiency of ready-made Expert Advisors with learning from the results of trading in the tester and have some developments, and at the moment I am just interested in your expertise, in terms of parsing, recognition, indexing for archives of different types of documents, albums, etc., if this post is not trolling, can look for mutually beneficial cooperation, contact me in private.

[Кеша Рутов]

Alexander_K:

Ehhhhh, guys...

Kesha has already become your savior... The grandson and faithful follower of SanSanych, who never even knew physics and mathematics...

Returns are basic, because the price is an integral part of them and nothing more.

The return is momentum, but there are also stochastic, mcdac, zigzag etc. Don't limit yourself, by God, some uncle, like "quantum from volstreet" says in the medium that returns are enough and everyone agrees with him, so quants do not use stoppers and tecrofits, for them it is all mathematics, it is another dimension, abstraction.

[Maxim Dmitrievsky]

Kesha Rutov:

The return is momentum, and there are also stochastic, makdak, zigzag, etc. No need to limit themselves, by God, some uncle, like "quantum from volstreet" in the medium that enough returns and all agree with him, well, quants and stoploops and teyrofit, for them everywhere one continuous mathematics, it is another dimension, an abstraction.

Aleshenka brother led us down the wrong road, with prediction of returns with negative error, and he himself ran away

[mytarmailS]

interesting question about predictive metrics was raisedhttps://stats.stackexchange.com/questions/126829/how-to-determine-forecastability-of-time-series

I don't know how important it is for our purposes, and how much it affects, but I didn't mind writing a few lines of code to check dependence of "some prediction power" on data window size.

So I took 4 different chunks of prices (returnees) and checked dependence of "prediction power" on the window size in each chunk

x1 is predictive power, x2 is number of data points in the window

Conclusions :

1) taking a fixed window when making predictions is far from optimal

2) the optimal window for forecasting is always "floating"

Code:

х <- my_price
q <- 10:500
ma <- matrix(nrow = 0,ncol = 2)
for(i in q) {
  
  x1 <- ForeCA::Omega(  tail(diff(x),i) ,spectrum.control = list(method = "wosa"))[1] 
  x2 <- i
  d <- cbind(x2,x1)
  ma <- rbind(ma,d)
}
plot(ma,t="l",main = max(ma[,2]))


где х  - это цена

[Yuriy Asaulenko]

mytarmailS:

interesting question about prediction metric raised

I don't know how important it is for our purposes, and how much it affects us, but I had the guts to write a few lines of code to check the dependency of "some predictive power" on the size of data window.

Conclusions :

1) taking a fixed window for predictions is far from optimal

2) the optimal forecast window is always "floating

Conclusions. More than'100 points forecast is meaningless.

[mytarmailS]

Yuriy Asaulenko:

Conclusions. It makes no sense to forecast more than'100 points.

No, the correct way to say it is pointless to take a fixed period