Machine learning in trading: theory, models, practice and algo-trading - page 2353

[Aleksey Nikolayev]

Maxim Dmitrievsky:

no way in forex)

Then if you go to other topics of the book, it becomes even more

Maybe to count separately for bids and asks, and then combine somehow? Most likely, it will turn out to be absurd.

It sounds logical, because it's already pretty clogged.)

[Mikhail Mishanin]

Aleksey Nikolayev:
Maybe instead of a hullabaloo we should do something more meaningful.) For example,take something from Prado apart. I found his idea of imbalance bars interesting but I have not understood how it can be adapted to forex.

Is there a Russian translation of Prado?

[Aleksey Nikolayev]

Mikhail Mishanin:

Is there a Russian translation of the Prado?

There is, but it is better in English - the narrative is brief and not simple, you will have to get the details in the articles, which no one will translate into Russian.

[Maxim Dmitrievsky]

Roman:

What is the point of his book then?

;))

there's useful stuff about resampling and random forest training, and in general it's a good material for getting acquainted with different methods

[Maxim Dmitrievsky]

Aleksey Nikolayev:

Maybe to count separately for bids and asks, and then somehow combine? Most likely, it would not make sense.

Sounds logical, because it's already quite polluted).

I don't know what kind of dream he had about such transformations but they only make sense when they actually have some sense) otherwise the same Renko

[Aleksey Nikolayev]

Maxim Dmitrievsky:

I don't know what dream he had about such transformations, but they only make sense when they really make sense ) otherwise the same Rencos

I don't know) But someone who wants to be like Prado, needs to think like Prado)

Yes, it looks like a Renko, but there are also some associations with CUSUM.

[mytarmailS]

How we can improve the predictability of time series

Using zigzag classification as an example.

Normalization by volatility

0) create an empty vector

1) follow the price in the sliding window of size n

2) normalize prices in the sliding window to the range 0-1

3) write the difference of the last normalized value with the previous one into the empty vector

4) make cumulative sum over vector

the code on P , at once with iterpolation of NA-boxes if they are present

roll.r01 <- function(x,n=10){
    res <- rep(0,length(x))
    for(i in n:length(x)){
      ii <- (i-(n-1)):i
      res[i] <- tail(diff(r01(x[ii])),1)
    }
    if(any(is.na(res))){
      print(   paste("WARNING vector haves NAs",sum(is.na(res)))    )
      res <- imputeTS::na_ma(res)
    }
    return(cumsum(res))}

auxiliary normalization function

r01 <- function(x)    (x-min(x))  /  ( max(x) - min(x))

This is what we get, the red row is the price, the blue row is normalized according to volatility

As we can see the series retains all the properties of the price but is more stable according to its characteristics

Let's try to compare the quality of the classification of the phase slope

the target - the declination of RG

signs - a dozen of standard indicators

AMO - forrest , with the same parameters and sids

trace 10k , test 10k

forecast at standard price

Confusion Matrix and Statistics

          Reference
Prediction   -1    1
        -1 3416 1894
        1  1582 3108
                                         
               Accuracy : 0.6524       

forecast at changed price

Confusion Matrix and Statistics

          Reference
Prediction   -1    1
        -1 3504 1568
        1  1332 3596
                                         
               Accuracy : 0.71           

I urge you to test!!!!!

[Andrei Trukhanovich]

Aleksey Nikolayev:

Are you hinting that it's time to leave your home nest of retail Forex?

There is a sense to stay if there is a small machine which can increase the deposit in a month, in all other cases it is easier to work in any other place.

[Forester]

mytarmailS:

How you can improve the predictability of time series

Using zigzag classification as an example.

Normalization by volatility

0) create an empty vector

1) follow the price in the sliding window of size n

2) normalize prices in the sliding window to the range 0-1

3) write the difference of the last normalized value with the previous one into the empty vector

4) make cumulative sum over vector

the code on P , at once with iterpolation of NA-boxes if they are present

auxiliary normalization function

This is what we get, the red row is the price, the blue row is normalized according to volatility

As we can see the series retains all the properties of the price but is more stable according to its characteristics

Let's try to compare the quality of the classification of the phase slope

the target - the declination of RG

signs - a dozen of standard indicators

AMO - forrest , with the same parameters and sids

trace 10k , test 10k

forecast at standard price

forecast at changed price

I urge you to test!!!!!

It is better to compare profits. Not an error in inclination.

[sibirqk]

mytarmailS:

How we can improve the predictability of time series

Using zigzag classification as an example.

Normalization by volatility

In fact, it's almost the same as building a trend line and then removing it from the original series. Yes, it is easier to predict this residual, but it all depends on the trend forecast. To forecast it we will need to know at least approximately where the price will go in the future. But if we know that, then why bother with all the previous stages?