Optimized Trend Trading | Page 28 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Optimized Risk vs Reward Equation 335 replies

Loss Less Optimized Strategy 63 replies

Optimized Parameters in Metatrader Testing? 3 replies

"Optimized Parameter" feature in EA? 0 replies

Strategy with Optimized parameter 1 reply

Attachments: Optimized Trend Trading
Exit Attachments

Optimized Trend Trading

  • #541
  • Edited 5:23pm Oct 22, 2009 5:11pm | Edited 5:23pm
Quoting Mechthildche
What does this imply if we do some kind of cycle analysis on the filtered data, then synthesize a signal using only cycles with a period >=2*105 bars and extrapolate it into the future? Would this extrapolated signal predict the future better than one based on data filtered with positive phase delay?
No, I don't think so. These filters are useless for smooting before spectral analysis because they're bad high frequency attenuators. Actually, for pre-smoothing I rather use the Discrete Cosine Transform.
  • #542
  • Oct 23, 2009 4:55pm Oct 23, 2009 4:55pm
Quoting asm8086
No, I don't think so. These filters are useless for smooting before spectral analysis because they're bad high frequency attenuators. Actually, for pre-smoothing I rather use the Discrete Cosine Transform.
Thanks for the excellent tutorial on ULLMA DF design. I'm curious how you use the DCT for smoothing. Do you simply set the high frequency components to 0 and invert the transform?

Do you have MT4 code for the DCT? I would be very interested in experimenting to compare the DCT with non-causal SSA as suggested by CB.

BTW. The Goertzel algorithm discussed so often on this thread does not require 3 cycles for correct computation. It is an IIR used as a FIR by cutting off the recursion after some period. Clearly it works for all frequencies within the analysis band, and only the lowest frequency satisfies the 3*period limitation. In fact the Goertzel will work for as many or as few cycles as you wish. I have modified the fixed block length Goertzel to operate as a variable resolution bank of filters, where each frequency is evaluated over a window that is a multiple of the period of the filter. It helps to decide which cycles are current. The only trick is to make sure that each filter is properly normalized.

I noticed in an earlier post by (I don't recall) that he could not obtain correct amplitude from his implementation of the Goertzel. His problem was normalization. The usual Goertzel must be normalized by a factor of 2/window length.

MC
  • #543
  • Oct 23, 2009 6:23pm Oct 23, 2009 6:23pm
Hi MadCow,
You are right, I set to zero all the bins below 8 or 16 bars and do the inverse transform. The DCT has a big advantage over the DFT because there's no Gibbs phenomenon caused by discontinuities, so the energy is well preserved throughout the window.

The DCT can be easily computed (here's an example for the 8 points DCT) :

0. Input signal : [1, 2, 3, 4, 5, 6, 7, 8, 9]
1. Symmetrical signal : [1, 2, 3, 4, 5, 6, 7, 8, 9, 8, 7, 6, 5, 4, 3, 2]
2. Apply DFT : [-N/8, -N/7, -N/6, -N/5, -N/4, -N/3, -N/2, -N/1, 0, N/1, N/2, N/3, N/4, N/5, N/6, N/7, N/8]
3. Thanks to symmetry the imaginary part cancels itself, so keep only the positive frequencies and you have a very simple DCT implementation.

Note that for a DCT of length N you must have N+1 points (at least in this implementation).

Philippe
  • #544
  • Oct 23, 2009 6:50pm Oct 23, 2009 6:50pm
Quoting asm8086
Hi MadCow,
You are right, I set to zero all the bins below 8 or 16 bars and do the inverse transform. The DCT has a big advantage over the DFT because there's no Gibbs phenomenon caused by discontinuities, so the energy is well preserved throughout the window.

The DCT can be easily computed (here's an example for the 8 points DCT) :

0. Input signal : [1, 2, 3, 4, 5, 6, 7, 8, 9]
1. Symmetrical signal : [1, 2, 3, 4, 5, 6, 7, 8, 9, 8, 7, 6, 5, 4, 3, 2]
2. Apply DFT : [-N/8, -N/7, -N/6, -N/5, -N/4, -N/3, -N/2, -N/1, 0, N/1, N/2, N/3, N/4, N/5, N/6,...
Thanks
MC
  • #545
  • Nov 4, 2009 1:02pm Nov 4, 2009 1:02pm
Quoting FxTimer
There is more to VFFT than it meets the eye,as usual,you have to do your homework,if you can access these 4 papers you will realize its possibilities :


1-Novel design of multiplier-less FFT processors -Yuan Zhou, J.M. Noras and S.J. Shepherd(Special emphasis on the phase-amplitude splitting technique that allows conversion to cyclic convolutions and additions.)

2-S. J. Shepherd, P. W. J. van Eetvelt, S. M. R. Jones, J. M. Noras and H.
S. Rajamani, "A Lower Complexity Discrete Fourier Transform",
Mathematics Today


Not so relevant,but...
What happened to FXTimer? Has anybody examined the concepts he refers to?
  • #546
  • Nov 4, 2009 1:17pm Nov 4, 2009 1:17pm
Quoting asm8086
I don't agree when CB said his filter's smoothing factor is equivalent to a 39 bar SMA. It's not because your filter's length is 39 the cutoff period is 39.
So we have to compare the CB's ullma type...
Looks similar to an analysis I did for myself, with the same result.
  • #547
  • Nov 4, 2009 3:53pm Nov 4, 2009 3:53pm
For the moment I didn't worked yet on this concept but I have the first paper : Novel design of multiplier-less FFT processors.
You can get it here :
Attached File(s)
File Type: pdf noveldesign.pdf   182 KB | 2,468 downloads
  • #548
  • Nov 4, 2009 4:02pm Nov 4, 2009 4:02pm
Quoting asm8086
For the moment I didn't worked yet on this concept but I have the first paper : Novel design of multiplier-less FFT processors.
You can get it here : Attachment
Thanks very much. I wasn't able to find either a copy of this paper, or the other papers referenced, though the book seems accessible enough.

Just wondering why this aspect of the topic has died down.

AugustLeo
  • #549
  • Edited 5:57pm Nov 4, 2009 5:47pm | Edited 5:57pm
BTW, I made a clone of the first CB's chart (post #1). You can download my Matlab code here :
Attached File(s)
File Type: zip TopCatFilters.zip   32 KB | 1,690 downloads
. Only run the Codebreaker.m function to get the following chart :

Attached Image (click to enlarge)
Click to Enlarge Name: topcat.gif Size: 20 KB


The algorithm :

1. Smooth the H1 candlesticks with HeikinAshi (ha)
2. compute the median price of ha ((hi+lo)/2)
3. Filter the median with the ullma type 0 (h1)
4. compute the trigger (predictive) MA (h2) by adding the momentum of h1 to itself

That's all...

Now I'm working on the clutter filter.

Philippe
  • #550
  • Nov 4, 2009 7:18pm Nov 4, 2009 7:18pm
Quoting asm8086
BTW, I made a clone of the first CB's chart (post #1). You can download my Matlab code here : Attachment . Only run the Codebreaker.m function to get the following chart :

Attachment

The algorithm :

1. Smooth the H1 candlesticks with HeikinAshi (ha)
2. compute the median price of ha ((hi+lo)/2)
3. Filter the median with the ullma type 0 (h1)
4. compute the trigger (predictive) MA (h2) by adding the momentum of h1 to itself

That's all...

Now I'm working on the clutter filter.

Philippe
So did you try simple crossover strategy on your MAs ?? If yes, can you post the results ??

Krzysztof
  • #551
  • Edited 4:05am Nov 5, 2009 3:32am | Edited 4:05am
fajst_k,

For now it's really useless to do some backtests... Clearly, if you trade with a simple crossover strategy you will always lose your money even with CB's MAs because some rules are missing, and if you're not convinced I cloned his MAs, think out my code, compare with the original chart and you'll see...
The signal to noise ratio or "clutter filter" like CB said is THE key of the TopCat system and I'm actually convinced that it's not computed by a spectral method but simply in the time domain.

Note that others methods CB spoke about too (like prediction/extrapolation by Goertzel, best timeframe finding by FNN, etc...) are subsequent work and for the moment, my aim is only to clone the TopCat system, so the rules we know are :

1. Trade EUR/USD on H1
2. Enter / Exit with ullmas crossover
3. Get confirmation with the clutter filter to avoid whispaw
4. Don't trade on weekend (see post #15)

So we now have half the system, and like I said I'm working on the other part : the clutter filter.
I'm very impatient to find the right method and then backtest the whole system.

Philippe
  • #552
  • Nov 5, 2009 4:20am Nov 5, 2009 4:20am
Hi,

pls find the 2 papers/articles CB referred to attached.
If needed I can also upload the less critical papers.

If anyone can get DCT in mq4 pls post.

cheers
Attached File(s)
File Type: zip VFFT2.zip   1.1 MB | 1,516 downloads
  • #553
  • Nov 5, 2009 4:22am Nov 5, 2009 4:22am
Quoting asm8086
fajst_k,

For now it's really useless to do some backtests... Clearly, if you trade with a simple crossover strategy you will always lose your money even with CB's MAs because some rules are missing, and if you're not convinced I cloned his MAs, think out my code, compare with the original chart and you'll see...
The signal to noise ratio or "clutter filter" like CB said is THE key of the TopCat system and I'm actually convinced that it's not computed by a spectral method but simply in the time domain.

Note that others methods CB spoke about...
Are you sure that what CB said here about trading systems is 100% correct ?? Because me not...See for example this - one of the J. Ehler eminiz strategies

8. Dual Super Smoother (DSS) Strategy

DSS is an acronym for a Dual Super Smoother. Super Smoother filters are described in Ehlers' book Cybernetic Analysis for Stocks and Futures. Two curves are smoothed with super smoother filters having different filter periods. Lag is removed from both filter lines and the crossing of the two lines are used as reversal signals in this strategy.



regarding clutter. See screenshoots. You can use optimizer to find thresold
values.

Krzysztof
Attached Image(s) (click to enlarge)
Click to Enlarge Name: chop1.JPG Size: 55 KB
Click to Enlarge Name: chop2.JPG Size: 91 KB
  • #554
  • Nov 5, 2009 4:30am Nov 5, 2009 4:30am
You're right fajst_k, what CB said isn't 100% correct and not so complex too, it's why I completely want to "reverse engineer" his system.
MisterH, thanks al lot for the article about VFFT, and if you have more, please send.

Cheers,
Philippe
  • #555
  • Nov 5, 2009 4:53am Nov 5, 2009 4:53am
Here is the complete article:
"A lower complexity discrete fourier transform"
It's from the British library, and they protect their content with a special reader.
Attached File(s)
File Type: zip A lower complexity discrete fourier transform .zip   1.2 MB | 1,231 downloads
  • #556
  • Nov 5, 2009 5:03am Nov 5, 2009 5:03am
Thanks a lot
  • #557
  • Nov 5, 2009 5:19am Nov 5, 2009 5:19am
after looking some more I found these 2 on my laptop:

"The Very Fast Fourier Transform ~ A Tool for Science in the 21st Century"

and

"Design and development of a high-speed winograd fast fourier transform processor board"

http://www.2shared.com/file/8950227/583e8cb7/VFFT3.html

note that CB said they are not as important as the first 2
  • #558
  • Nov 5, 2009 11:37am Nov 5, 2009 11:37am
Quoting asm8086
... I'm actually convinced that it's not computed by a spectral method but simply in the time domain ... So we now have half the system, and like I said I'm working on the other part : the clutter filter.
Phillipe,

Why do you think that Codebreaker's clutter filter is not computed via a spectral method. In this post (#195 in the thread) he indicates that he does use a spectral method.
Quoting Codebreaker
I use the ULLMA to get a smooth price. I then do a spectral analysis based on the Goertzel algorithm and assign dominant cycles to "signal" and smaller amplitude cycles to "noise". Hence we get SNR. This helps to keep the system out of fake trades and whipsaws.
I got the impression that this post was referring to his "TopCat" system.

AugustLeo
  • #559
  • Nov 5, 2009 12:53pm Nov 5, 2009 12:53pm
Quoting AugustLeo
Phillipe,

Why do you think that Codebreaker's clutter filter is not computed via a spectral method. In this post (#195 in the thread) he indicates that he does use a spectral method.


I got the impression that this post was referring to his "TopCat" system.

AugustLeo
That post was quite a bit later than TopCat description. He did not have the Goertzel algorithm when he first described topcat. Some later posts discuss the coding of the algorithm.

He is defining it this way because he thinks the dominant cycle is the trend, everything else is noise. But using this definition and the Goertzel, you will not get the rapid response shown in the original topcat image because a Goertzel, or any other DFT will not change rapidly enough, since it is an average over a long block. I think his later systems may use the Goertzel definition, but not topcat.

I agree that his original clutter filter must depend on price action. How else to get such rapid response?
  • #560
  • Nov 5, 2009 12:55pm Nov 5, 2009 12:55pm
Does anyone have the algorithm for performing a Bartels test on cycles?

Thanks.

AugustLeo
1 trader viewing now