Adaptive digital filters - page 2

[Prival]

Well, he was wrong to go after Kalman. If I understand correctly, it is talking about Kalman filter with constant coefficients known as alpha-betta-gamma filter (these are different modifications of Kalman filter).

You need a Neutron.

here we have compared Kalman filter (correct) and Butterworth filter.'Random Flow Theory and FOREX'.

There's an algorithm for calculating them on matcad. If anyone ventures to make a Butterworth filter in MQL I can help (explain what's there and how it's calculated in Matkadec), and I don't think JMA will be better (you can compare).

Kalman is inherently iterative MNC so there's only one way around it, if the models embedded in the filter don't match the process under study. (They just don't know how to prepare it :-))

Understand the word adaptation implies an answer to the question of what you have to adapt to. In radiolocation there are such concepts as signal (useful component) and noise (what disturbs us). Having understood this question it is possible to make adaptive filters, until you answer this question it is not clear what you need to adapt to.

[Сергей]

NightPaul:

2 grash
Here is what the author himself writes about JMA ) - http://www.jurikres.com/catalog/ms_ama.htm#top
Since all this is for sale, we get only disassembled code, as I understand it, and I really want to understand what is the trick )

Thanks for the link. I think this ferret is using a rather tricky adaptive filtering algorithm, most likely with prediction elements based just on autocorrelation. I think so. :o)

to Prival

Understand the word adaptation implies an answer to the question of what to adapt to. In radar there are such concepts as signal (useful component) and noise (what disturbs us). Having understood this question it is possible to make adaptive filters, until you answer this question it is not clear what you need to adapt to.

How true it is, Prival ... I agree completely. By the way, I remember that you and your colleagues tried to build some ACF indicators and started with the "ACF Indicator". I have had an "observation" here.

[Prival]

To grasn ACF was made, and if it weren't for the help of forum members, I would still be messing around. And so a lot of good and useful things have been done, at least from my point of view. I've stopped now, I need to think, a lot of things are falling apart because of one thing, not knowing the sampling rate of the process. I think I will start a new topic tomorrow, I need to get current estimates of price and sampling rate with a minimum confidence interval. Then move on. So far I am writing an introductory page to explain what it is for, to make it clearer. If it is possible about it a little bit more "I have some "observation". Someone who knows the words "Widrow-Hopf" can have very interesting observations IHMO

[Сергей]

to Prival

It doesn't matter what I researched, but I 'caught' the ACF for company. This is just an observation, not confirmed by anything, roughly speaking, looked at the results with an eye and "caught it". Not proven by anything, not statistically confirmed, probably complete nonsense, but worth checking on occasion. The point is to make some assumptions about the development of the series by the type of ACF. So far I have roughly classified 2 variants (ACF is taken from the black series, gray series - process development). I am giving it without special comments, it is kind of obvious:

Variant A

Variant B

PS:

One who knows such words "Widow-Hopf" may be very interesting ...

Prival - I have written, that in DSP - I am self-taught and my limited and technical illiteracy is obviously not enough to understand, that Nyquist frequency rules the world ...

[Prival]

I do not know what you can see, I can see from ACF that option A can predict for 200 counts (I do not know what you have on the X-axis minutes or something else). Option B is 50, then the nature of the process changes, but you need to look at the dynamics, because the ACF changes over time. And the first thing this function shows is the correlation time (time during which the process can be predicted) + the second is the type of process itself, almost always an oscillatory circuit (in radio engineering terms); we can further classify it by types of oscillatory circuits but in my studies (at this stage) it is not so important. At first it is necessary to deal with one kind of an oscillating link, with others by analogy it will be easier.

[Сергей]

Prival:
I don't know what you can see, I can see from ACF that for variant A I can predict for 200 samples (I don't know if you have X minutes or what else). Option B is 50, then the nature of the process changes, but you need to look at the dynamics, because the ACF changes over time. And the first thing this function shows is the correlation time (time during which the process can be predicted) + the second is the type of process itself, almost always an oscillatory circuit (in radio engineering terms); we can further classify it by types of oscillatory circuits but in my studies (at this stage) it is not so important. At first it is necessary to deal with one kind of vibrating element, it will be easier to deal with others by analogy.

I tried to classify "by species and type" by simple observations:

• For the "type" of ACF of variant A: the process most often did not move away from its average values and kept its statistics for some time.
• For the "type" of ACF of variant B: the process more often "fell off" from its average values.

In general - if it's not the main thing, then it doesn't matter... good luck

[Sceptic Philozoff]

Here's what the author himself writes about JMA. http://www.jurikres.com/catalog/ms_ama.htm#top

This JMA is very impressive, very impressive. I somehow didn't pay too much attention to it before, as I have a preconceived notion of muwings. But now it looks like I'll have to reconsider it.

As for that JMA that's in Code Base ('JMA'), it clearly doesn't look like the original one. Yes, it's smooth, but it clearly lags more.Parabellum's drawing there is much more convincing.

And there again appears the problem I am struggling with: I want to transform the chart of initial quotes so that it could eliminate disasters and then apply Jurik's indictors (or their clones) to the transformed chart... Somehow it seems to me that even if the distribution turns into something similar to Gaussian, nevertheless the price process will not be Wiener-like - because its Hurst index will be more than 0.5 (because of dependence of neighbouring samples).

P.S. Prival, again to you: http://www.jurikres.com/faq/faq_ama.htm#betterthan . Especially look at the third figure from the bottom: JMA, unlike other filters, has practically no Gibbs effect (spike after gap). And there are effective techniques to remove this effect (when I was a student I came across a book by Hemming "Digital Filters", I need to find it).

[Rashid Umarov]

Mathemat:

Here's what the author himself writes about JMA ) - http://www.jurikres.com/catalog/ms_ama.htm#top

It's a very impressive JMA, very impressive. I somehow didn't pay too much attention to it before, as I have a preconceived notion of muwings. But now it looks like I'll have to reconsider it.

As for that JMA that's in Code Base ('JMA'), it clearly doesn't look like the original one. Yes, it's smooth, but it clearly lags more. Parabellum's drawing there is much more convincing.

Here is information for reflection on one Jurik indicator - J_TPO

[Prival]

Here's the hemming

[Sceptic Philozoff]

Prival, thanks for the book. And here's another surprise for you, confirming your view of price as a target:

Conquering lag while making no simplifying assumptions (e.g., that data consists of superimposed cycles, daily price changes having a Gaussian distribution, all prices are equally important, etc.) is not a trivial task. In the end, JMA had to based on the same technology the military uses to track moving objects in the air using nothing more than their noisy radar. JMA sees the price time series as a noisy image of a moving target (the underlying smooth price) and tries to estimate the location of the real target (smooth price). The proprietary mathematics is modified to take into account the special properties of a financial time series.

Taken from the same place, emphasis added.

Second. JMA doesn't redraw, so there's no FFT to speak of. Nevertheless they have removed the Gibbs effect...

Third. The Jurik Research team assumes something similar to the Cauchy distribution as a distribution model. What it is, you know: none of the moments of this distribution exists, not even m.o. Feel the ambush the enemy has set for us? Although, on the other hand, it's possible that their goal was simply to build an inductor to effectively smooth out even random wanderings with Cauchy-distributed increments.

2 Rosh: well, at least you solved the mystery of one Jurik inductor. Respect.