I don't have it yet! So I thought I'd ask your advice on whether it's worth getting involved in all this. Who's thinking about it?
I don't have it yet! So I thought I'd ask your advice on whether or not to get involved in all this. Who's thinking about it?
Can we get on with the party.
Seryoga hi! (I left a message for you in a neighbouring thread, hope you saw it :)
ALL SIMPLER!!!!! My idea was quite simple - and the hands did not really get around to it. There's no need to run anything here and there. A perfect approximation of the low frequency component would be:
[1/window*(SUM history per window)+ 1/window*(SUM , future values per window)/2
So it all comes down to predicting an average of some fixed window. And it can be predicted by autocorrelation methods. I'm 100% sure it will work much more reliably and accurately. Think of it as a miniature adaptive filter.
It can be improved to the "mind"
to Vinin
Give us your thoughts - there will be a sequel!
to grasn
Hi Sergey!
Of course I did. It was them (your considerations) that led me to this idea of non-harmonic synthesis. Unfortunately all my efforts to predict by autocorrelation methods crashed at the event horizon. We need non-linear correlation methods with elements of adaptation.
to grasn
Hi, Sergei!
Of course I've read it. They (your considerations) are what led me to this idea of non-harmonic fusion. Unfortunately all my efforts to predict by autocorrelation methods crashed to the event horizon. What's needed here are non-linear correlation methods with elements of adaptation.
No, no, no, no. If you reconstruct the time series by the predicted mean, it won't work, large errors. We need not that, we need to evaluate the local extremums of the predicted "ideal LF curve", but it's actually a pivot zone!!!! You should be less demanding :o)
I would see how the weights behave on history. That is, I would make an indicator with three buffers: w1,w2 and w3.
No problem. Only what will it give us? Clearly they will behave regularly with a fluctuation period of a smaller scale, since they are the solution to a cubic equation.
No, no, no, no. If you reconstruct the time series by the predicted mean, it will not work, large errors. We need not that, we need to estimate the local extremums of the predicted "ideal LF curve", and these are actually pivot zones!!!! One should be less demanding :o)
Here I don't get it!
There are no stable relationships there.
No, no, no, no. If you reconstruct the time series by the predicted mean, it will not work, large errors. We need not that, we need to estimate the local extremums of the predicted "ideal LF curve", and these are actually pivot zones!!!! One should be less demanding :o)
Here I don't get it!
There are no stable relationships there.
And I didn't get your non-harmonic analysis either. Who is the perfect derivative and where does it come from? If you explain the essence of your analysis step by step, I would be grateful (I think I'm not the only one).
In the meantime, I'll mess around in Matkadec and in a couple of days or sooner (I'm all about statistics) I'll give you a detailed description of my idea :o)
PS: maybe their integration will yield something interesting :o)
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
You agree to website policy and terms of use
Here's a thought that excites me.
Let's take the most common wizard with an N averaging window. Let's run it through the time series (RT) forward and backward, thus eliminating group and phase delay and obtaining an ideal smoothed curve, the first derivative of which optimally shows the entry and exit points... although only on historical data.
It is connected with inevitable overrendering at the right edge of the data, the further into the history we recede, the less this effect will affect us. In the limit, it can be neglected at a distance N from the leading edge. Thus we have the task to predict this derivative for N bars ahead (fig. on the left).
We can do it in another way. Let's perform only straight run of the mask, we will get a standard smoothing, which we all got used to long ago, with N/2 lag (fig. on the right). The task can be set as prediction of derivative values for N/2 bars ahead. By the way, both in the left fig. and in the right one N is chosen so that the LPF bandwidth would be roughly the same - 100 bars for a two-pass scheme (left) and 200 bars for a straight run (right). So, we'll have to make equal forecast for the same number of bars ahead, but the derivative is smoother for the double-pass scheme, which means better prediction accuracy.
I should say right away that all attempts to forecast by "usual" methods will not give a positive result, as soon as we get closer to event horizon (N/2 or N) the forecast accuracy rapidly goes down bringing it to zero at the very horizon. Such is the fundamental property of BP...
So I was thinking, what if for a given BP I'll build a fan of one-pass mashups with step 1 starting with N=2 or even 1 and up to 1000 for example. It is clear that the informativity of adjacent swipes is not very different, so let's construct an autocorrelation function that shows the "similarity" of adjacent swipes (or their derivatives). As would be expected, a number of consecutive swabs are highly correlated (Figure left):
Since the informativeness from correlated instruments is low, we will thin out the series of swabs and leave only those whose correlation coefficient between them does not exceed 20%. There are only three of them left - with an averaging window of 6, 80 and 300 bars. We now take and equate the weighted sum of values of the lagging bars to the ideal derivative (the red thick line in Fig. right): dMA=w1*dMA1+w2*dMA2+w3*dMA3.
We should construct three such equations for three consecutive bars on the right side of history minus N/2 (to avoid chattering), solve them in relation to weights w and calculate the value of dMA on the right side of BP. That's it! We obtain the prediction value, which indicates the expected direction of BP.
A kind of non-harmonic analysis is obtained :-)