Optimized Trend Trading | Page 33 | 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

  • #641
  • Dec 10, 2009 1:22am Dec 10, 2009 1:22am
Quoting fajst_k
...

So long TF traders watch out - Random Walk is behind you

Krzysztof
Krzysztof, although not having read all of the study yet, I doubt that (practically over the long run) smaller timeframes are more profitable. There must be other factors to consider than Hurst Exponent...

Quoting MisterH
...

The fact that the H1 mentioned in this thread is favored by some -including me- is associated with a certain study of a major investment bank from a couple of years ago on Mutual Information and timeframes showing that it does not make any sense to trade a time frame smaller than 60 minutes.
...
  • #642
  • Dec 10, 2009 4:32am Dec 10, 2009 4:32am
Quoting enigmaFX
Krzysztof, although not having read all of the study yet, I doubt that (practically over the long run) smaller timeframes are more profitable. There must be other factors to consider than Hurst Exponent...
I think they are simply because of this basic rule.

Or put it in another way. If you take let's say 4H TF than to achieve the same measurement error like in 1M TF you need 240 times longer time. Than you will trade much less frequent and make less money.

Krzysztof
Attached Image (click to enlarge)
Click to Enlarge Name: error.JPG Size: 129 KB
  • #643
  • Dec 10, 2009 12:18pm Dec 10, 2009 12:18pm
I tend to agree with enigma, in too small TF the spread becomes an issue.
  • #644
  • Dec 10, 2009 12:44pm Dec 10, 2009 12:44pm
I think by now it's pretty clear that DFT, Goertzel, ... are pretty much useless on fx time series. The problem is chronomorphicity. It's a difficult word for saying that things change over time: volatility, autocorrelation, fractal dimension, ... but also the dominant cycle length. Who cares what the dominant cycle was over the past 180 bars. The only thing what matters (to me as a trader) is what its length is in the current bar. Because that's where I have to take my decision.

What I have found to be a good way for extracting the dominant cycle length of a financial time series is an indicator John Ehlers introduced: it's called the "homodyne discriminator". It comes from the Hilbert(-Huang) Transform: those who have read "Rocket science for traders" by Ehlers will be familiar with it.

If you take this dominant cycle (which changes all the time) according to the homodyne indicator, divide it by 2 and use this as an input for your indicators, they all become adaptive. Just try it: they become much smoother and more responsive to price action, compared to giving them a certain value (like 14 is often given). You simply "listen" to the market instead of "imposing" your own idea of what the cycle length (and hence the length of your indicator) should be. Some indicators need half a cycle, others (like CCI) need 1 full cycle as input.

Now the most simple indicator you can find is momentum: difference between the close of the current bar and the close of a bar N periods ago. N in this case will be half the dominant cycle length.

Now my question to you is:
1) does anyone think he can find a smoother that is -almost- as good as the yellow line (it's a SMA10, shift-3). Universal approximators (Neural network, fuzy logic, machine learning...) or some kind of DSP trick, it doesn't matter, I'm kind of desperate...

2) does anyone have experience with adjusting the length of the smoothing filter? So the number of periods to smooth over is also dynamic?
if so, please pm me.

If we can smooth the histogram, we have an adaptive leading indicator that is a little like Martin Pring's KST, but faster!
Attached Image
Attached File(s)
File Type: zip FF.zip   1 KB | 1,326 downloads
  • #645
  • Dec 10, 2009 1:03pm Dec 10, 2009 1:03pm
Quoting MisterH
1) does anyone think he can find a smoother that is -almost- as good as the yellow line (it's a SMA10, shift-3). Universal approximators (Neural network, fuzy logic, machine learning...) or some kind of DSP trick, it doesn't matter, I'm kind of desperate...
http://en.wikipedia.org/wiki/Smoothing_spline
The breaking of a wave cannot explain the whole sea.
  • #646
  • Dec 10, 2009 1:17pm Dec 10, 2009 1:17pm
Quote
What I have found to be a good way for extracting the dominant cycle length of a financial time series is an indicator John Ehlers introduced: it's called the "homodyne discriminator". It comes from the Hilbert(-Huang) Transform: those who have read "Rocket science for traders" by Ehlers will be familiar with it.


I believe you are not so much updated. John gave up with frequency discriminators in 2007 (Hilbert transform based) see screenchoot. Homodyne discriminator has a lag around 20 bars !!! Just make a simple strategy using MAMA and you will see how easy is to loose money.

Krzysztof
Attached Image (click to enlarge)
Click to Enlarge Name: zzz.JPG Size: 69 KB
  • #647
  • Edited 3:14pm Dec 10, 2009 2:32pm | Edited 3:14pm
Quoting fajst_k
John gave up with frequency discriminators in 2007 (Hilbert transform based) see screenchoot. Homodyne discriminator has a lag around 20 bars !!! Just make a simple strategy using MAMA and you will see how easy is to loose money.
Krzysztof

On page 121 of his book, "Cybernetic Analysis for Stocks and Futures" Ehlers states:
"The lag of measuring the dominant cycle period is about eight bars." In 2004 He called it the most responsive technique available. So if Ehlers gave up on frequency discriminators in 2007; what did he replace them with?

update: here's a link to his entire presentation;
http://www.mesasoftware.com/Seminars...sInTrading.ppt
  • #648
  • Dec 10, 2009 3:10pm Dec 10, 2009 3:10pm
with strategies from from eminiz and corona charts

presentation is

http://www.eminiz.com/in/Presentations.aspx

title Colleagues in Trading 2006

Krzysztof
  • #649
  • Edited 5:44pm Dec 10, 2009 5:31pm | Edited 5:44pm
Hello,
I want to clarify something...

Quoting MisterH
What I have found to be a good way for extracting the dominant cycle length of a financial time series is an indicator John Ehlers introduced: it's called the "homodyne discriminator". It comes from the Hilbert(-Huang) Transform: those who have read "Rocket science for traders" by Ehlers will be familiar with it.
John Ehlers work don't have nothing to see with the Norden E. Huang's work (from Nasa).

Huang invented an adaptive and data driven process named Empirical Modal Decomposition (EMD) to decompose non-linear, non-stationnary multi-components signals into simples AM / FM components and then apply the Hilbert transform on them to get the instantaneous spectrum (some improvements are possible because the hilbert transform isn't really appropriate for this job). The whole process is called HHT (Hilbert-Huang Transform).

John Ehlers implements a 4th order FIR Hilbert transform in a recursive way to get out of the frequency/magnitude uncertainty limitation, but it lags about 8 samples (because of too many smoothing) and it works ONLY for "pure" sinewaves. I want to say this principle is MATHEMATICALLY INCORECT when applied on complex (ie multi-components) signals ! So the homodyne discriminator always supply fake frequency estimation when applied to financial data. Be careful with Ehlers work.

Quoting MisterH
I think by now it's pretty clear that DFT, Goertzel, ... are pretty much useless on fx time series. The problem is chronomorphicity.
About the Goertzel :
Perhaps the magnitude information has a 3/2N phase delay but the phase information has zero lag (it's normal because Goertzel is a bandpass filterbank and all IIR bandpass filters have zero lag at central frequency) so there is a way to get something really usefull from it, even for trading.

Philippe
  • #650
  • Dec 10, 2009 7:37pm Dec 10, 2009 7:37pm
Quoting MisterH
On page 121 of his book, "Cybernetic Analysis for Stocks and Futures" Ehlers states:
"The lag of measuring the dominant cycle period is about eight bars." In 2004 He called it the most responsive technique available. So if Ehlers gave up on frequency discriminators in 2007; what did he replace them with?

update: here's a link to his entire presentation;
http://www.mesasoftware.com/Seminars...sInTrading.ppt
Thats not homodyne discriminator but another method of cycle
measurement based on Hilbert transform. Homodyne was descibed in Rocket Science....lag 20.5 bars...anyway both methods in my opinion dont work well..

Krzysztof
Attached Image (click to enlarge)
Click to Enlarge Name: HD.JPG Size: 144 KB
  • #651
  • Dec 11, 2009 12:55am Dec 11, 2009 12:55am
Quoting fajst_k
I think they are simply because of this basic rule.

Or put it in another way. If you take let's say 4H TF than to achieve the same measurement error like in 1M TF you need 240 times longer time. Than you will trade much less frequent and make less money.

Krzysztof
Krzysztof, certainly higher sample rates have smaller measurement errors and so with lower timeframes we get a more accurate picture of the market data. But what is the market data? It consists of signal and noise. Our goal is to find a STABLE signal (cycle) for "predicting" the market future. I very rarely find them in timeframes below 1h (most likely I cant see them because of too much noise or these smaller cycles are just not stable long enough).

So basically what we achieve with a higher sample rate is to more accurately sample the noise part of the data, but for the relevant signal portions (> 50 cycle on 1h) this lower measurement error is absolutely irrelevant.
  • #652
  • Dec 11, 2009 7:07am Dec 11, 2009 7:07am
Quoting enigmaFX
Krzysztof, certainly higher sample rates have smaller measurement errors and so with lower timeframes we get a more accurate picture of the market data. But what is the market data? It consists of signal and noise. Our goal is to find a STABLE signal (cycle) for "predicting" the market future. I very rarely find them in timeframes below 1h (most likely I cant see them because of too much noise or these smaller cycles are just not stable long enough).

So basically what we achieve with a higher sample rate is to more accurately sample the noise...
Are you able to confirm this with strategy results with a few hundreds trades ??

There is a lot of opinions like this but never confirmed somehow....

Krzysztof
  • #653
  • Dec 11, 2009 9:33am Dec 11, 2009 9:33am
Quoting fajst_k
Are you able to confirm this with strategy results with a few hundreds trades ??

There is a lot of opinions like this but never confirmed somehow....

Krzysztof
What do you want me to show you: that money can only be made on higher TFs, but will be lost when trading on smaller TFs? "It's the strategy, stupid!" - money can be made (and lost) on all TFs.

I just can tell you that MY trading strategy (partly based on DF) works consistently well on larger TFs, but is doing miserable on smaller TFs. This is still manual trading (otherwise I wouldn't be here but sailing through the caribbean islands together with Megan Fox).

Have a look at the attached screenshots (1h, 15min and 1min, all same settings, not optimized for any period, taken randomly): on the 1h chart you can see that the peak of the STLM cycle corresponds quite well with the peak of the price trend (I've marked this with a vertical line). On the 15min this is less frequent, on the 1min we almost never hit the price peak. So which TF do you think gives the best "signal" for this strategy?

BTW: on which TF could you identify more stable (predictable) cycles?
Attached Image(s) (click to enlarge)
Click to Enlarge Name: h1.JPG Size: 122 KB
Click to Enlarge Name: m15.JPG Size: 110 KB
Click to Enlarge Name: m1.JPG Size: 93 KB
  • #654
  • Dec 11, 2009 10:06am Dec 11, 2009 10:06am
Quoting enigmaFX
What do you want me to show you: that money can only be made on higher TFs, but will be lost when trading on smaller TFs? "It's the strategy, stupid!" - money can be made (and lost) on all TFs.

I just can tell you that MY trading strategy (partly based on DF) works consistently well on larger TFs, but is doing miserable on smaller TFs. This is still manual trading (otherwise I wouldn't be here but sailing through the caribbean islands together with Megan Fox).

Have a look at the attached screenshots (1h, 15min and 1min, all same settings, not...
test is very simple: the same data set (length) than strategy optimized on 1min than optimized against the same fitness function on higher TF. Than compare results: Profit Factor and Sharpe ratio for both of the strategies. In both strategies must be like > 200 trades each to have statistical signifiance of this test....than maybe you can say something if those tests will be consistent for different currency pairs.

BTW. i see you use STLM...perhaps generated by digital filters program.
Did you ever consider lag of the filters generated by this program. If not
you can put coeffcients of this filter to MATLAB and make a printout. You can be very suprised i think.....

Krzysztof
  • #655
  • Dec 11, 2009 10:30am Dec 11, 2009 10:30am
Take a look at Ehlers' corona cycle period vs. his homodyne discriminator. There's quite a difference...
Attached Image (click to enlarge)
Click to Enlarge Name: coronavshomodyne.jpg Size: 33 KB
  • #656
  • Dec 11, 2009 10:56am Dec 11, 2009 10:56am
Here is Hurst exponent calculated for 1min and 4h TF for different lengths. It is
calculated with 'Servicik" method which I think is one of the most accurate. Referencial Random Walk
0.5 in blue.

It's clear from this that when the length and TF rises than we go away
from anti-persistant regime to more supper-diffusive regime (trending).

So it looks that for higher TFs more trend oriented strategies should be used at least for this particular case.

Krzysztof
Attached Image(s) (click to enlarge)
Click to Enlarge Name: SE_HE.JPG Size: 167 KB
Click to Enlarge Name: SE_HE1.JPG Size: 189 KB
  • #657
  • Dec 12, 2009 4:09am Dec 12, 2009 4:09am
don't forget that if you start with a small TF like 1M for example, you can always resample it as you see fit and try to find optimum resampling time to maximise your SnR. It may lead to the conclusion that the 1H or 4H is the best, but it may lead to 20 min or 100 min.

When you start with a higher TF there is some information you lost forever.

My 2 pips
  • #658
  • Dec 12, 2009 5:15pm Dec 12, 2009 5:15pm
Quoting challenger78
don't forget that if you start with a small TF like 1M for example, you can always resample it as you see fit and try to find optimum resampling time to maximise your SnR. It may lead to the conclusion that the 1H or 4H is the best, but it may lead to 20 min or 100 min.

When you start with a higher TF there is some information you lost forever.

My 2 pips
You are absolutely right. The higher TF is just a subsampled lower TF, so info has been lost, and noise added in the form of high frequency aliasing. You can get the same spectral information from the lower TF, however most of the indicators s.a. Goertzel don't work when the period is stretched too far (the algorithm works, but the implementation is limited). e.g. to get a 50 hourly period analysis using M1 data you need a window that's at least 50*3*3600. Since the aliased noise is all pretty high frequency, the subsamping process does not really increase noise viewed through the Goertzel filters, except at the high end.

But why hassle over the correct TF? Everyone has opinions. Does anyone have facts.. like a system that works at one TF and not another?
  • #659
  • Dec 12, 2009 7:46pm Dec 12, 2009 7:46pm
Quoting MadCow
You are absolutely right. The higher TF is just a subsampled lower TF, so info has been lost, and noise added in the form of high frequency aliasing. You can get the same spectral information from the lower TF, however most of the indicators s.a. Goertzel don't work when the period is stretched too far (the algorithm works, but the implementation is limited). e.g. to get a 50 hourly period analysis using M1 data you need a window that's at least 50*3*3600. Since the aliased noise is all pretty high frequency, the subsamping process does not really...
We're not even at the system point here. In this tread we went from promising potential to mostly disappointment. BUT some of us here are taking the right approach - Characterizing our data, and only then moving on to find tools/filters to help build a system.

So we're quite a ways from building a system - we're starting to characterize our data, but still need to find tools that help us identify current properties of our data (trends, whipsaws, sideways, etc).
Once we've got those tools we can go on and build various systems from the tools in our toolbox. But first we need to get those tools.

The hassle over TF is important as we need to provide good data to our tools. Garbage in, garbage out as they say.
  • #660
  • Dec 13, 2009 7:46am Dec 13, 2009 7:46am
Quoting fajst_k
So it looks that for higher TFs more trend oriented strategies should be used at least for this particular case.
Krzysztof
or in other words: a trend oriented strategy is supposed to work "better" on higher TFs - 100% agreed.

Anyway, enough of theory and TFs - I stick with flyboy that we should start building some practical useful tools.

I am using 2, sometimes 3 periods (xTLM indicators) as a kind of clutter filter to give me the trading direction. I am using FIXED periods - in my opinion there is no need to overly optimize the filter to a certain HISTORICAL period. I am interested in the FUTURE trend, hence I need filter(s) flexible enough to help me detect stable periods as early as possible. Have a look on H1, H4 or even D1 TFs of e.g. EURUSD or GBPJPY. The cylces of about 50 periods come and go - sometimes they stay only for several bars, sometimes they are there for weeks. But sooner or later they will come!

I am searching for these periods/TFs where the market is (more or less) "in sync" with my filter (basically when we see more or less something related to a sine wave). Then I start trading my strategy on this TF.

I am a trader and not a math expert, but perhaps some of you with better math skills could come up with ideas on building a clutter filter based on this.
5 traders viewing now