Fast Fourier Transform - Cycle Extraction - page 9

 

Goertzel output

Simba said:

"I think that we have been misreading Goertzel indicator..instead of using ,for example,0.050 as 5% of 720 periods(36 bars cycle)...I believe that the right reading is 0.050 periods per bar(20 bars cycle)...I have done a test with values 360,540 and 720 and the results achieved have led me to believe this is the right interpretation"

So I understand it is correct interpretation of output from Goertzel ??

You said your indicator is composite. But what is a weigth of different components ?? Equal ??? Because it should be not equal I think.

We have a different TFs and different players with different money power

are there, their systems start to play at the bar close (1M,5M, 15M etc)

but with different money power so cycle gets different strength 'kicks' at those closes. If you added all frequencies extracted from all TF with equal weight thats wrong. Example of this can be MTF stochastic oscillator strategies where stochs must synchronize in direction but different weigths are given to different TFs stochastics --- see FF stochs threads.

Krzysztof

 

Yes. The output of Goertzel is 1/periods gives you the cycle length. And it seems that you cannot just pick and choose which frequencies you want to use (that was what I thought). All of the frequencies must be used in combination due to the dynamics of the additive sine waves of different amplitudes and phase shifts.

Your weights are the amplitudes of each frequency. Each frequency (sine wave) will add or subtract at different times depending upon the phase shift as well... if that is what you are referring to.

As far as the MTF idea, what I hope to end up with is a composite, dynamically "weighted" waveform for each available timeframe. And, as far as "weighting" each timeframe, I won't know how to proceed until I see the results from my research.

It has been interesting to see the waveform change as I add the shorter wavelengths.

Carl

fajst_k:
Simba said:

"I think that we have been misreading Goertzel indicator..instead of using ,for example,0.050 as 5% of 720 periods(36 bars cycle)...I believe that the right reading is 0.050 periods per bar(20 bars cycle)...I have done a test with values 360,540 and 720 and the results achieved have led me to believe this is the right interpretation"

So I understand it is correct interpretation of output from Goertzel ??

You said your indicator is composite. But what is a weigth of different components ?? Equal ??? Because it should be not equal I think.

We have a different TFs and different players with different money power

are there, their systems start to play at the bar close (1M,5M, 15M etc)

but with different money power so cycle gets different strength 'kicks' at those closes. If you added all frequencies extracted from all TF with equal weight thats wrong. Example of this can be MTF stochastic oscillator strategies where stochs must synchronize in direction but different weigths are given to different TFs stochastics --- see FF stochs threads.

Krzysztof
 

Meyers documents

Hi,

Did you read documents from Meyers Analitics about Optimization, Goertzel&Fourier than endpointFFT ?? He explains there why iFFT fails and shows optimization methods which makes FFT strategy profitable. I think

this optimizatiom method should be used to find proper "weigthing" as "weigthing" is changing in time, and this dynamic weigthing which you use will be just "curve fitting"........past values adjusted for maks profit will not work in future. I think it is necessary use extrapolation for this.

Do you know what is the status of Simbas and Barnix research using SSA and MESA together using SSA-MTM kit ?? They dont post anymore about it.

I believe they wanted to get better resolution or better confirmations of peaks. Did they acheve something ??

Regarding SSA it is non casual (useless for future prediction) but according to this post from FF Noxa made casual version based on SSA and NN:

Hi Noxa !

I have several questions about CSSA :

1. If I understood, the first two steps (buid the trajectory matrix and compute the SVD) are both identical in SSA and CSSA, so I guess the causal trick takes place only at the reconstruction time, or I'm wrong ? ie, the singular spectrum from SSA and the singular spectrum from CSSA are both the same ?

2. If you sum every outputed causal components at each instant n, does the signal will be perfectly reconstructed (equals to the original) ?

3. About the smoothing parameter, does it serves for pre or post processing ?

Thanks a lot in advance for your informations

Phil

Have you ever thought that a causal version of CSSA could be learned from SSA with a recurrent neural network such as an Echo State Network?

Noxa

Do you have file #_lib_FFT.ex4 maybe ?? iFFT indi use it.

Krzysztof

 

Yes, the Goertzel flattens price before applying its wave analysis.

I have read some information about how Brian Millard analyzes. So, essentially I have a long way to go.

Now, since I can extract frequencies, I need to create a moving window and perform a Goertzel for each window, and record which waves occur. I then move the window forward and repeat until all data has been analysed. Then I need to find which (4) four frequencies occur consistently in the past, preferable a high, median term, and low. Those will have a higher probability of continuing to occur in the future. Then I can perform future projections. I can analyze all timeframes to see which one has the most consistent cycles.

Carl

 

MTF analysis

Hi,

I attach some basics from DSP book. From this equation you can see that an error depends from number of samples so if you take e.g. 1 week data 4H

we will have 240 times bigger error than analyzing 1 week data 1min.

Beside this I believe that 4h, 1d data contains much bigger error simply because 4h or 1d candle closes in different time in different time zones.

so 4H 'cycle kick' takes place in different moments. I write this because I've seen in other threads a lot of analysis on 4H and D1 TF.

1M data contains everything but it very noisy....

Krzysztof

Files:
error.bmp  896 kb
 

Stationary - non stationary

Here is another basic thing. Because Forex data is a mix of stationary and non

stationary data (trend and choppy market) analyzing of this data will be very difficult with poor results. Much better is just cut the signal. See below.

So did you try this MTM toolkit ?? Did Barnix and Simba achieve anything practical ??

Krzysztof

Files:
 

Here's another example of a signal, where a cross is combined with a "dip".

Also find attached the extrapolator template I made.

The folder with screen-captures was too large to upload to this forum, so you can download it from here.

Enjoy!

 
poruchik:
Extrapolator

Extrapolator - MQL4 Code Base

Method 1:use Quinn-Fernandes Algorithm

Method 2: Autocorrelation Method

Method 3: Weighted Burg Method

Method 4: Burg Method with Helme-Nikias weighting function

Method 5: Itakura-Saito (geometric) method

Method 6: Modified covariance method

Thanks for posting this indicator! I find it very interesting and useful.

crodzilla:
More repainting indicators? We don't need any more of those...

Yes, it does repaint quite heavily. But different than with other repainting indis, here, depending on the settings and the way you read it, it appears to give signals that are nevertheless most reliable.

The past few days I've been fooling around with the 2nd mode, which is based on a autocorelation calculation that seems to be used particularly in audio technology.

In the setting that works for me I use a couple of curves, where the "LastBar" settings are respectively 13, 21, 34, 55, 89. Any set of curves with shifted LastBar will probably do fine, but this set is based on Fibonacci sequence.

The trick is NOT to simply take a signal from a curve that seems to indicate a trend reversal. Instead, only look for crosses, or "dips" that occur for what seems to be a new top or bottom.

For an example of reliable crosses, see picts 1 and 2.

For an example of a "dip", see the pict 4, and the example of the next post I'll make (due to restriction in max attachments).

The 5th pict shows how prices evolved further along the way.

For a reliable long/short entry on a dip, it should looke like entering on the bottom/top of a B-wave of an ascending/descending Elliot-wave type A-B-C cycle (or of what can potentially be the 2nd wave of a 5-part cycle).

This seems to work over and over. Although there's certainly a lot of repainting, the signal will be reliable on the critical candle itself, and will not give a false signal too soon. You need a bit of experience to get a good feel for it and recognize the set-ups correctly.

Note how on the 3rd chart you can see a cross, but you can also see that a new bottom is indicated. This means that to go long you have to wait until you reach that new bottom, or until it is gone due to the repainting. Generally, these dips don't get wiped out by the repainting, though, and will remain identifyable until entry.

Any of the extrapolators can show a dip that can be used for an entry.

As it is, presently I like to use it to confirm and finetune my entries. If you have a good set-up or general method, then I think this indicator can really help in gaining reliability, and can keep you out of trades when you happen to make some kind of error (especially critical on fast timeframes).

In the next post, I'll have attached a folder with screen-captures of the EUR-USD 4 hr of a couple of recent critical reversals. Just unpack the folder, and browse back and forth through the charts with your picture viewer, so you can get an idea how the indicator repaints and can be used reliably.

You will also find attached the template I made for this.

It's not really clear to me why this indicator has to repaint. It is supposed to work with a fairly straightforward autocorrelation calculation that should not really have to repaint. So maybe there are some improvements possible here...

Enjoy

 

hi

hi samurai

thanks for sharing i give it a look

regards lodol

 

Sigview

It look that this thread is dying a bit. Nobody intrested in cycle extraction ??

You have a cycle == you have a proper setting of digital filter or stochastic

or RSI and than you know when the trend turns.... Here is a bit o fuel. Below

analysis of EURUSD using sigview program (demo obtainable from sigview site).

It is very powerfull tool, it has e.g. different functions to deal with input signal (window filters) and also signal calculator which allows to combine singnals. Have a look.

Krzysztof

Files:
sigview.jpg  134 kb