Trading Strategies Based On Digital Filters - page 37
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
lost one frequency for clear signal
GOLD15 main signal sin 0.5HZ + cos 0.1HZ -- SA didn't find slower frequency for 600 bars but it foud both frequencies for 200 and 400 bars
lost resolution for signal with noise/ faulty peaks
GOLD30 - main signal with smoothed noise It created two clear peaks for 600bars but for 400 and 200 bars it started to show additional peaks so it lost resolution significantly.
disaster for pure noise !!!!
GOLD60 gauss noise smothed - disaster !!! shows different peaks with variable amplitude depend of number of bars. Less bars ==> higher peaks.....
next disaster for mixed signal
GOLD240 - mixed signal, first noise than signal + noise. Next disaster, different peaks dependent of number of bars.
It was very interesting to follow this thread from the beginning till end. DF generator program together with MESA SA build in, some articles showing that is works etc, etc. But during the reading, perhaps because of my proffesion, (for several years I was testing, finding and fixing faults in
large telecom software systems), my thought was: Where is the proper test of this system ?
It can not be done on FOREX data as this data has unkown structure which this system should find. It must be done on fictious data with known structure to discover this structure first.
Than when I reached the end of thread I asked SIMBA for conclusions but no answer somehow
https://www.mql5.com/en/forum/175938/page21
than I decided to make test myself.
For this I generated following fictious data (attached .hst files) and transferred it to MT.
GOLD240- 300 bars of gauss noise smothed with 15SMA + 300bars of 0501sincos signal with gauss noise sm with 15SMA
GOLD60 - 600 bars of gauss noise smothed with 15SMA
GOLD30 - 600 bars of 0501sincos signal with gauss noise sm with 15SMA
GOLD15 - 600 bars of 0.5HZ sin + 0.1HZ cos signal
GOLD5 - 600 bars of 0501sincos signal with gauss noise
GOLD1 - 600 bars of gauss noise
Than I applied build MESA SSA from DFG program first as it is input for generating DF, I knew what I should get. I did this test for 200, 400 and 600 bars. Later on I did those tests for SA from MTM toolkit with GRACE.
Unfortunatelly results were not astonishing.
GOLD15 main signal sin 0.5HZ + cos 0.1HZ -- SA didn't find slower frequency for 600 bars but it foud both frequencies for 200 and 400 bars
GOLD30 - main signal with smoothed noise It created two clear peaks for 600bars but for 400 and 200 bars it started to show additional peaks so it lost resolution significantly.
GOLD60 gauss noise smothed - disaster !!! shows different peaks with variable amplitude depend of number of bars. Less bars ==> higher peaks.....
GOLD240 - mixed signal, first noise than signal + noise. Next disaster, different peaks dependent of number of bars.
CONCLUSIONS.
SA only recognised clear signal (GOLD15) even in this case failed once for 600bars !!!!. It lost resolution very fast for signal with noise and for clear noise and mixed signal it has shown faulty peaks. So this system can only be used for a data series when we are sure that they are not mixed with random data and S/N ratio is high enough. See pictures below. I hope those test will help you.
KrzysztofHi,Krzysztof
I just became aware now that you had posted such questions,so I will summarize my reply.
1-MESA is not very good for noisy data,so,either we use it with a S/N filter,like Damiani`s volatimeter or we use it on smoothed data or we expose ourselves to nasty surprises.
2-Goertzel is better for noisy data,not perfect ,but a good tool,so,you may want to use it instead of MESA.
3-I prefer real time band pass filters to JMA,SATL,FATL,T3,Astrology,etc...if you want to use a low pass filter any of the above will do.
4-Conman disappeared?no idea,I don`t use it anymore,it was done with DFG and Mesa,I believe the ex4 should be at the thread.
5-Cycles are non stationary,though some of them have a certain persistence(high bartels),so we can use them .
6-Dickey fuller example was at the FF thread,if it is still there(it should be since I didn`t delete it) just download it and take a look,if it isn`t there,you will have to do it by yourself like clahn and myself did just reading CB explanations.I don`t have it anymore nor do I have the ConMan since I systematically delete everything I am fairly sure I won`t use again.
7-The Question:What is the cause of Cycles?
Regards
Simba
fajst_k, I have been analyzed your data with Matlab's pmcov() function and have got the same results for false peaks. May be 200-400 bars is insufficient for correct analysis?
If you want, I can post here my screenshots.
You need to Digital Filter Generator software: https://www.mql5.com/en/forum/172930
Thanks, I have that software, by the way one can download it directly from the author's home page, in a single file.
http://fx.qrz.ru/ this web page is the English translation of the original web page in Russian.
But I still need help on how to use the spectrum cycle peak in an EA code.
reading peaks
I think you should read this thread from the beginning. In earlier posts Simba
explains very detailed how to read those peaks than later I show that SA
fails pretty easy to show good peak against noisy and random data.
The only way is to do it manullly against pieces of data which we are sure is
not random and not very noisy.
Krzytsztof
Damiani Volatmeter test
Hi Simba,
You wrote
1-MESA is not very good for noisy data,so,either we use it with a S/N filter,like Damiani`s volatimeter or we use it on smoothed data or we expose ourselves to nasty surprises.
Than I made a test of Damiani Volatmeter. I applied to it gauss noise so it should show no signal. See below. It shows total b.s. a lot of green signal
above grey.
I checked the code and this what it shows is
ATR(1) STD(1)
------- - -------
ATR(2) STD(2)
So kind of change of range or volatility but you dont know if it is because
of chage of signal amplitude or noise amplitude....so it has nothing with S/N ratio.
If you still have Dickey-Fuller document on your PC can you post it here. It disappeared from the link in FF (and the excel sheet)
Krzysztof