Machine learning in trading: theory, models, practice and algo-trading - page 3186

New comment
 
fxsaber #:

Started resizing (within reason for scalping) the min knee of the ZZ and watching the sum of the knees.

What's a reasonable range for scalping? From... to?

I have on bars from 0.00200 for EURUSD, something begins to earn. But I'm afraid that there is the adjustment you described above. I.e. the best variants on OOS, which are spoilt simply by shifting the window for training by 2-10% (i.e. 2-10% of lines are different, as a result trees are built differently and OOS is completely different, up to unprofitable).

 
fxsaber #:

ZЫ In general, if there is interest to try to find differences between the two rows, can provide them.

Rather interested in a real training dataset for your data (fit and target). To see how much I could make on scalping with my approach.

 
fxsaber #:

Answer yourself the question what useful characteristic is retained/lost. With my conversion I answered such a question.

If I understand your transformation correctly

you take the increments, then randomly select the index of the increment and either leave it as it is or invert it (x/-1).


Let's take some abstract series with some obvious structure.

apply your transformation.

No structure is preserved, it's just random from random...

I wouldn't consider it a simulation at all...



Here's the code.

par(mar=c(2,2,2,2),mfrow=c(3,2))

r <- rep( c(1:10,10:1) ,3)
r |> plot(t="l",main = "начальный ряд с какой то закономерностью")

rd <- c(0,diff(r))

for(i in 1:5) {
  sa <- sample(1:length(rd),size = length(rd)/2)
  rd[sa] <- rd[sa]/-1
  rd |> cumsum() |> plot(t="l", main = "ваше преобразование")
}
 
Forester #:

Rather interested in a real training dataset for your data (fic and target). To see how much I could make on scalping with my approach.

Choose a broker for historical data according to the highest potential profit. For example, EURUSD_Broker1 has a higher potential profit than EURUSD_Broker2. Then take EURUSD_Broker1.

Both majors and crosses can be scalped. But not all of them. Just train on each of them and see the results. That's what I do, roughly speaking.

 
mytarmailS #:

If I understand your conversion correctly

you take the increments, then randomly select the index of the increment and either leave it as it is or flip (x/-1)

You have understood my transformation absolutely correctly.

Let's take some abstract series with some obvious structure (regularity).

The transformation is so simple that even without graphs it is clear that you can get anything up to a monotonically increasing price.
 
fxsaber #:

You have understood my transformation absolutely correctly.

The transformation is so simple that even without graphs it is clear that you can get anything up to monotonically increasing price.

Well, then it can't be considered a simulation,

and you have to realise that you've destroyed all the trends that were there,

as well as the very structure of the DEM.

 
mytarmailS #:

Well, then it can't be considered a simulation,

and you have to realise that you've bucked all the trends that were there,

as well as the very structure of the DEM.

I'm sure I destroyed far less than the alternatives. However, it was enough.

As for the structure, this method is based on the law of large numbers, on which the structure is based. It is exactly for cases where there are tens of millions of original data.


I don't want to praise myself, but I highly doubt that there is such work with ticks and such powerful invariants: time, spread, absolute increment (as a consequence - fat tails, non-stationarity, correlations, etc.). I.e., there are a hundred million input data invariants. No comparison to the "100" stat characteristics of the models.


Despite such unique features of randomisation, a test was immediately found that unmistakably answers the question of what we are dealing with: randomisation or reality.


There is nothing to argue or debate here. What is demonstrated is not just a difference between SB and reality, but a subtle difference between the original series and randomisation. The value is in the counterexample.

 
fxsaber #:

Looks like I got an interesting Random generation.


Cool idea! I just need to figure out how to use it 😆 In essence, it turns out to be a price BP with the same session characteristics, the same volatility as the real one, but stationary. With constant expectation mate and apparently constant variance. And the tails are Gaussian. At the same time, the profit on average is zero minus the total spread.

 
fxsaber #:

I'm sure I destroyed far less than the alternatives. However, it was enough.

I think you overestimate this method, maybe I'm missing something.

 
I didn't understand the meaning of what happened at all, but this is a normal state of mind for me.
1...317931803181318231833184318531863187318831893190319131923193...3512
New comment
Reason: