Machine learning in trading: theory, models, practice and algo-trading - page 2316

 
Valeriy Yastremskiy:

sophisms are all ))))) Without intermediate knowledge, there will be no knowledge)) at the intermediate stage is anything, the SB is usually like times of change))))

got it

 
Maxim Dmitrievsky:

only I have not yet understood which button to press to get better results for the classification

only an example for the regression

i know that the default value of sampling with the maximum gradient (this is a new feature) is used

or it is just built-in by default and you don't have to do anything

By the way catbust is very tough in terms of retraining... it's very hard to make it retrain. If the dataset is crap... ...it will learn badly and not remember all the options.

Apparently we are fixing at different stages of the process) I'm interested in the process of building the initial probabilistic model - if it more or less describes reality then any more or less sensible algorithm will give results in subsequent calculations.

 
Aleksey Nikolayev:

Apparently, you and I are fixed at different stages of the process) I am interested in the process of building the initial probabilistic model - if it more or less describes reality, then any more or less meaningful algorithm will give results in subsequent calculations.

Apparently there are not many who can understand how a proper probabilistic model can describe reality. ))) Any movement in the models?

I sometimes run a minority game model. It often looks like a tick chart at times.

 
Valeriy Yastremskiy:

There don't seem to be many people who can understand how a proper probabilistic model can describe reality. ))) Any movement in the models?

Sometimes I scroll through the minority game model. It often looks like a tick chart at times.

I don't see much (practical) point in global price models that describe it in all aspects and over all time. It seems to me that only models that describe individual aspects of price on separate time intervals can be useful. As a simple example I can cite my article about gaps, which examines only the maximum deviation of the price to the opposite side of the gap, before its closing (separate aspect), and respectively, only on the time interval from the gap formation to its closing.

Not that this is a new idea) Rather, just stating the obvious out loud.

 
Aleksey Nikolayev:

I don't see much (practical) sense in global price models that describe it in all aspects and over all time periods. In my opinion, only models that describe individual aspects of price on individual time intervals can be useful. As a simple example I can cite my article about gaps, which examines only the maximum deviation of the price to the opposite side of the gap, before its closing (separate aspect), and respectively, only on the time interval from the gap formation to its closing.

Not that this is some new idea) Rather, just stating the obvious out loud.

model and analysis of systems with a large number of elements can only start with the obvious sections)

I've never understood why or what the hep is closing with. I take it as a non-obvious fact)

 
mytarmailS:

I had the following idea

1) Teach the neural network some actions, let it be buy/sell.

2) the network will make a lot of mistakes on the new data

3) I wanted to cluster patterns in layers of the network to see if it is possible to separate erroneous decisions of the network from correct ones by looking at patterns that appear in the network during signal processing...

=========

Vladimir, do you know if there is a package in R-ka in which you can interact with the graphics, for example, so that I can select an area on a graph with my mouse and get the parameters of this area in the code

I haven't. I can't even imagine how it can be done.

 
Valeriy Yastremskiy:

I've never understood why and what closes a hep. I accept it as not an obvious fact.)

It closes with a price lower than the gap (when the gap is up, for example). What else can it close with)

If the price after the gap = SB, its closing will occur with a unit probability and with a known distribution for the maximum deviation of the price in the other direction.

 
Vladimir Perervenko:

I haven't. I can't even imagine how it can be done.

It can be done with the locator() function, run it with the right parameters and click on the chart...


Here's an example of how to build a trend line by clicking on a chart with a mouse

data <- matrix(cbind(1:50,cumsum(rnorm(50))), ncol = 2)
plot(data,t="o")

xy <- locator(n=2)
xy$x <- round(xy$x,0)

abline(coef(lm(y~x, xy)),col=8)
lines(xy, col="red", lwd=2)

I wrote a function that selects rectangles on a chart

rect.locator <- function(){ 
  L <- locator(n=4)  
  id <- c(1, which.min(abs( L$x[1] - L$x)[-1])+1)
  idd <- c(mean(L$x[id]),mean(L$x[id]),mean(L$x[-id]),mean(L$x[-id]))  
  q <-  L$y[id][1]
  idy <- c(1, which.min(abs( q - L$y)[-1])+1)
  val <- c(mean(L$y[idy]),mean(L$y[-idy]),mean(L$y[idy]),mean(L$y[-idy])) 
  return(list(idx=idd,val=val))}

x <- cumsum(rnorm(100))
plot(x,t="l",col=1)

r <- rect.locator()

rect(r$idx[1],xleft = r$idx[3],ytop = r$val[1],ybottom = r$val[2],
       col= rgb(0,1,0,alpha=0.3))


You can teach the neuronku such things that you don't know how to program, but we can draw them and after getting the parameters of the drawn can send them in the form of images to the neuronku, I think it's very cool))

Or with locator() we can easily select some specific clusters, or ... or ... anything you like, dream on....


Here's code how to select interesting parts of the chart and apperceive with furtively

x <- cumsum(rnorm(200))
plot(x,t="l",col=8)

 xy <- locator(n=2)
id <- round(xy$x,0)
idd <- id[1]:id[2]

library(dtt)
dc <- dct(x[idd])
dc[7:length(dc)] <- 0
dc <- dct(dc,inverted = T)

NAve <- rep(NA,length(x))
NAve[idd] <- dc

lines(NAve,lwd=2,col=4) 
abline(v=range(idd),col=8,lty=2)


 
Aleksey Nikolayev:

It closes at a price below the gap (at a gap upwards, for example). What else can it close with)

If the price after the gap = SB, then its closing will occur with a unit probability and with the known distribution for the maximum deviation of the price in the other direction.

The SB is clear. But it closes rather quickly, sometimes it's even possible to make profit))) although it may happen.

 

Forum on trading, automated trading systems and trading strategies testing

MetaTrader 5 Strategy Tester: bugs, bugs, suggestions for improvement

fxsaber, 2021.01.31 11:06

Very often in the analysis of TS (machine learning models) we look at Sample and Out of Sample on the chart of single passage.

To understand where the OOS is on such a graph, you have to hover with the mouse and look for the date in the tooltips.


After that you begin to understand where this OOS is.

And so every time you have to move the mouse, look and so on.


I propose to add a function.

void TesterGraphLabel( const datetime LabelTime, const string Description ); // На графике одиночного прохода ставится соответствующая метка.

With saving the label in the tst-file.


This will greatly simplify the work with the OOS and other cases.

Please comment on the proposal. I may be missing something or seeing it narrowly. Comment not here, but there.