Durschner Moving Average 3.0 - Wealth Lab Forum

Carova

#1

1/14/2018 1:20 PM

Hi Cone / Eugene!

Is there any possibility getting this (http://www.vtad.de/sites/files/forschung/M_Duerschner_Gleitende_Durchschnnitte_3.pdf) translated into a WL Indicator? I don't read German and my attempts to use Google Translate to help me understand the concept has been a disaster. (Google Translate does not do well for "technical" documents I have found).

Vince

Eugene

#2

1/14/2018 1:43 PM

Hi Vince,

I'm puzzled. If you don't read German how can you understand the concept? And if you don't, what makes you interested in it? If there's some code in a language like Metastock, Excel or EasyLanguage then it might be possible to understand it, though.

Carova

#3

1/14/2018 2:02 PM

Hi Eugene!

That is exactly the problem, I do not read German so I am having trouble understanding the concept. However the attached graphic (from http://www.scientific-trading.com/avg3g.php) leads me to believe that it may have significant value.

Vince

Carova

#4

1/14/2018 2:08 PM

graphic

MA3.0.PNG

Eugene

#5

1/15/2018 9:40 AM

Vince, here's English version of Durschner's article from a TA journal:

IFTA Journal 2012 (PDF), see pp. 27-31

Hope this helps.

Carova

#6

1/15/2018 10:48 AM

Thanks Eugene!

This helps a lot.

Vince

DrKoch

#7

1/15/2018 5:46 PM

Hi Carova,
I don't want to disappoint you, but this article (I am german) is quite a lot of hokuspokus and the mentioned "Nyquist Theorem" is not applied correctly.

If you are interested in "better" moving averages, you should indeed think about low-pass filter theory. This theory is quite old. Results for "optimal" filters (with different criteria what is meant by "optimal") have the names "Butterworth", "Bessel" or "Tchebychew".

I once coded quite a complete set of theses filters in the old WL world. Probably you find a port to WL6 somewhere.

Another approach is the Kalman filter. Also a well funded theory.

In old WL there used to be a script called "MA contest" which shows a practical comparison of all the moving averages which stem from these ideas.

Carova

#8

1/15/2018 5:54 PM

Hi Dr. Koch!

Welcome back! Thanks for the advice. I never studied filter theory in school (I was not an EE) so I never could understand how to calculate the coefficients required from the wavelengths. :(

I sure do wish that the old WL4 forum had not been removed by the WL folks. It had an enormous amount of very useful material (such as the MA Contest script that you wrote) and thoughtful discussions in the forum that I sure wish I could still mine.

Would you happen to have a copy of the WL4 MA Contest script handy?

Vince

Eugene

#9

1/15/2018 7:04 PM

Typo:

QUOTE:
"Tchebychew".

Should be:
Chebyshev filter

superticker

#10

1/15/2018 11:08 PM

QUOTE:
I don't want to disappoint you, but ...

What WL users should appreciate here is the way the financial world reports "daily bars" is not the way it would be done in the engineering world. The way your CD player works is the "better" approach. It's Nyquist frequency is at 22KHz so data are sampled on the CD at 44KHz. Then in a leap a faith (Let's not go there.), the CD player algorithm "tries" to oversample that to twice or triple that and then runs it through a 6-pole comb filter to give it a sharper cutoff frequency at 22KHz. If you have a $400 CD player, look at its spec sheet for details.

By the same token, an engineer would send the 30-minute bar data (in effect, oversampled) through a low-pass Bessel filter to compute the daily bars. This gives a sharper cutoff with minimal phase delay and sampling artifact. (Wealth Lab could be enhanced to do that [say with a special oversampling option], but it would be the only trading platform sporting that feature. Wild! Wealth Lab could redefine the financial world.)

So the way daily bars are actually determined in the financial world--without oversampling--creates "some" sampling artifact (as determined by the sinc function). Is this a problem? Well, the people that wrote the cited paper seem to think so. https://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem

My gut feeling is that there's so much more random noise in this fuzzy system problem that a little high-frequency sampling artifact isn't going to make that much difference. Obviously, the authors of the paper would disagree. I would be interested in comparing the results with and without the 3rd generation moving average to see if removing some of the sampling artifact actually helps. As a control, it would be nice to also compare it to the normal case where oversampled data is used to compute the daily bars with a low-pass zero-phase filter.

Size:

Color:

[quote]I don't want to disappoint you, but ...[/quote]What WL users should appreciate here is the way the financial world reports "daily bars" is not the way it would be done in the engineering world. The way your CD player works is the "better" approach. It's Nyquist frequency is at 22KHz so data are sampled on the CD at 44KHz. Then in a leap a faith (Let's not go there.), the CD player algorithm "tries" to oversample that to twice or triple that and then runs it through a 6-pole comb filter to give it a sharper cutoff frequency at 22KHz. If you have a $400 CD player, look at its spec sheet for details.

By the same token, an engineer would send the 30-minute bar data (in effect, oversampled) through a low-pass Bessel filter to compute the daily bars.  This gives a sharper cutoff with minimal phase delay and sampling artifact. (Wealth Lab could be enhanced to do that [say with a special oversampling option], but it would be the only trading platform sporting that feature. Wild!  Wealth Lab could redefine the financial world.)

So the way daily bars are actually determined in the financial world--without oversampling--creates "some" sampling artifact (as determined by the sinc function). Is this a problem? Well, the people that wrote the cited paper seem to think so.  [link]https://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem[/link]

My gut feeling is that there's so much more random noise in this fuzzy system problem that a little high-frequency sampling artifact isn't going to make that much difference. Obviously, the authors of the paper would disagree. I would be interested in comparing the results with and without the 3rd generation moving average to see if removing some of the sampling artifact actually helps. As a [u]control[/u], it would be nice to also compare it to the normal case where oversampled data is used to compute the daily bars with a low-pass zero-phase filter.