Optimized Trend Trading

that's nothing unusal for human traders scalpers - see this

http://www.trade2win.com/boards/meta...ml#post1059760

Its equity curve of my friend who is scalping in his shop > 100 trades from demo account. technique is very simple EMA 5,21,100,200, stoch MACD +
watching 20 charts at the same time.

However he is not able to explain his decisions, not able to explain why he is winning or losing. So in certain moment he can start to loose and he will also not know why.

Have you read 'Evidence based technical analysis' book ?? There are some chapters about human traders.

So mixing technology like 'Lapunov Exponent' with human intervention is not the best idea i think becouse your mind makes biased decisions.

Krzysztof

Thank you a lot of the reference.
Of course when I ask friends to use even simple statistical tools to test their setups I have no answer.

Even the simple chi-square test. But no answer.
In fact I use the fractal break - out setup explained for the first time several years ago on the TSD.

In fact this helps me to identify similar conditions. For example for a Ross hook after a fractal break - out I have found statistical significant results.
I try to explain to friends that the best way is to have several clearly defined patterns tested on a particular, frame, time and pair.

But the traditional wannabe technical analysts use this form of analyses as a menu of patterns. And that is WRONG. Also wrong is to transpose some techniques from the stock market into forex without any tests.

For example you see a pattern on a 4 h frame.
You see the same pattern on 15 m frame. Those patterns are different. The second may be just a product of the cyclic volatility, and the first one have a different meaning.
You see a pattern on a 1 m chart. This one can be just the result on the algorithm of the retail broker. It can be a stochastic event too.

The form is not enough.

Hello!!

Some interesting indicators...

Best regards.

And this one claims to predict S&P using LagLasso regression and Kalman filter

http://videolectures.net/amlcf09_mahler_mtsp/

easy to check

Thank you all very much.
55 gbpusd can you tell if you are able to run the posted indicators some posts ago?
And can you explain what are for your indicators.


Krzysztof, thank you for the article. I had a look to your recommended book. There is missing the fractal brownian motion hypothesis.

Hello!!

I discovered yesterday these indicators in this russian forum

http://forextechnologies.ru/for/


I attach a picture comparing STO JMA ( more smoothing ) with a normal stochastic.

As usually I cannot answer the question WHY, but I can answer the question HOW.

I think I found a reason. In fact this indicator do not load under some metatraders. For example the same indicator I cannot make it work in Alpari. But I can make it work in FXOpen. Exacltly the same indicator, strange. indeed. Open a demo account and try.

I just do not know why. But I found HOW.

Second, I really find interest here in digital filters for newbies as myself.

http://depositfiles.com/ru/files/3084863

This is a link I found some time ago in the TSD forum, and it is still working. Inside you can find a plethora of digital filters created by the best russian hackers. Everything is in just one place.

Some people may hate me for that, maybe we should start a new tread, digital filters for newbies, or optimized trend trading for newbies and migrate.

Maybe, some dark TSD members ??

Don´t worry, I have not seen anyone to be profitable with DSP in this forum.

I attach a picture of a good smoothed filter with a minimum lag. ( not repaints )

If anyone are interested, pm please, I will share for free..

For me this thread has been an absolute waste of time. Nothing really beats a JMA to smooth a time series (I mean the Tradestation version from Jurik Research). I challenge anyone who reads this to show a better filter than the JMA.

So what is your exact judgement method ?? How you are measuring the error ??

Tibetan plateau, Everest, digital filters and other things

Nice challenge. But what would be its use?

OK, you have the JMA in Tradestation, a great filter and a priori you have paid for it. Knowing a bit about human psychology it will be almost impossible to prove to you that there is any better digital filter. But it is not necessary and I will try to argument why.

You have a great filter. But after that you ask yourself what would be the settings 9, 12, 15?

What would be the phase -100, -50,5,50,+100.

And every different settings are giving a different solution.

Of course Jurik's filters and Ehler's filters are state of the art. And of course some talented coders and hackers try to reverse engineer and to make even better.

Another point is that even different algorithms under some conditions will give the same result with minor differences. Which is better, this one or that one. Sometimes one is better sometimes the other is better. And you are in the quest of the best algorithms. And when you have the "best" you start to wonder what settings to use?

The digital filters by themselves do not give an edge. On the contrary they have a dark side.

Understanding what they are doing and in what market conditions they operate best is a possible answer.

In this forum I made a little bit of hypothesis for the use of the fractal dimension indicator. And this is public and open source. For example in the official mql forum there is another indicator IVAR, estimating the local Hurst Exponent. There is an article unfortunately in Russian, so a few are able to understand it.

Behind the highly complex mathematics the basics are: the price time series have two main states. They are either:

persistent (there is a bigger probability that the next movement will be in the same direction than the previous)or

anti-persistent (there is a bigger probability that the next movement will be in the opposite direction than the previous)

the 0.5 Hurst exponent is an exception and not a rule

Without understanding that operating a digital filter may be a dangerous exercise. The digital filters are much more reactive and they will react when a SMA, will not react. But under some conditions it will be wrong to take the signal. And as a rule every wannabe digital filter power user try to obtain the best reactivity. And he faces the reality, that even the best digital filters are not crystal balls.


So my hypothesis is as follows:
Digital filter in their trend following algorithms form operate best when the price time series are persistent.

Digital filter in their oscillator algorithms operate best when the price time series are anti-persistent.

And basically there are two ways of optimization:
Neural nets:
Genetic optimizers:

I think that the optimization problem is much more important than the filter problem. How and When? When to trade and when not to trade?

The digital filters with Neural Nets have inherent dangers.
In fact the digital filters are offering much more options to explore to the Neural nets than an ordinary indicator.
Doing that there is a bigger risk of over fitting. In fact that is the biggest danger of the neural nets. They are looking for a perfect but local solution and the digital filters are always offering that kind of local solutions. As a friend of mine said" I am looking not for Everest, I am looking for the Tibetan plateau of the good solutions.
http://articles.mql4.com/479

Genetic Optimizers and NNs are only a small subset of different Machine Learning tools. NNs are looking for local solutions but e.g. SVM no. All depends from learnig algorithm. NNs usally are using gradient descent which is out of date a bit. SVM are using QP, SMO or LS method.

Thank you for the clarification.

In fact I mean that those two machine learning tools are most popular among traders.

There are many different optimization algorithms for the Neural Nets.
Anyway the professionals make their own code and implement it. For the newbies like me we have to use open source software.

As a free software an interesting choice is Rapidminer.
You can find there SVM, but the Neural Net Models are disappointing.

http://www.neuralmarkettrends.com/tutorials/
Here you can find a tutorial how to use the Rapidminer.

Recenlty I found another interesting software developed primarily to support particle interaction classification in high energy and neutrino physics experiments at CERN.

http://www.ire.pw.edu.pl/~rsulej/NetMaker/

For the sake of completeness (if rapidminer is mentioned one must also mention its competitor knime):

http://www.knime.org/

Both are quite interesting and their feature lists are of the same caliber. Some things work better in rapidminer while other things are better implemented in knime.

Thank you for the reference of knime. They have some very interesting neural net stuff, that is completely missing in rapidminer they have only multilayer perceptron.

Nevertheless, there are interesting things undermetatrader.

I attach a picture. This is a very famous system a reverse engineered asctrend. However I added to the stop line a jma smoothing.

The signals are with risk=8: in fact this is the recommended setting.

For a signal it is necessary a confirmation between the stop line and the signal line.

I have plotted a SSA of bands as a general indication of the trend. It is not necessary but it is cool, to see the position of the price related to the bands.

Below I have plotted a Normalized SSA and a PFE with digital filter. So This makes me a really nice oscillator. The Normalized SSA is giving signals for a possible change in momentum but they are not trading signals.

Below there is a measure of the strength of the trend. You should know this indicator it is a digital analogue of ADX.

And far below I have the FGDI. I like this version of the FDI, because it has a standard deviation channel. For me ot is important because, the fractal dimesion and the Hurst exponent are not calculated, they are estimated and there are many different ways of doing that.


Blue FGDI: Range probability and Pink noise phenomena
Red FGDI: Trend probability and Black noise phenomena

All the indicators above the FGDI are measuring different components of the price time series. Only the FGDI is giving us probabilities what is going to happen in the future.

And the break - out from a zone of blue FGDI to a red FGDI there is an increased probability of technical break - out.

It looks like Disneyland but behind there are some very interesting algorithms.

And what happened next. The fractal break - out made a warning for a reversal or new movement. We know that something is going to happen but we do not know in what direction.

Observe there is another fractal break -out.

I think it is clear. It is useless to give endless examples.

Shouldn't the SSA cycle and the green trendline change prior to 16:00pm? I compared your chart with the chart you posted on post#835 and SSA did not seem to repaint as it should. Does your chart post #835 used future data or did you use the Noxa CSSA instead?

Of course the green line repaints. But I use it as a general guidance where the trend is going, i like it. And when it repaints instead of complaining I know that this has a reason and I start to look at the fractal dimension structure.
(For example if we are with Low fractal dimension (red FGDI) and that happens, it is likely that this was provoked by black noise, expect a powerful reversal)

As for the normalize version I plot it together with a sound indicator PFE.
I want to know how it changes vis-a-vis this indicator.
NO the post did not use future data. I do not know if it is possible under MT4, I used lag bars for the SSA calculations

Repainting/Recalculation is a phenomenon. Why it repaints. Because it uses future data for the calculations. To minimize this fact I use a lag, and that is the simplest but not the optimum solution. another approach is to tune the SSA to the market conditions or to use a neural net (or something else) to predict the future bars and make a calculation based on that, but if you can predict the future bars why should you use SSA LOL/There are guys out there who claim the whole approach of SSA is proven useless and Independent Component Analysis or Wavelets instead would be better.

When and what causes a repainting if a lagged SSA is the real question?

In this case the fractal break - out made a prediction that it is likely something is going to happen. If the break - out is in the direction of the SSA, there would be no repainting, if it reverses it will repaint. And there is 50 - 50 chances, as far as I know.


I made today some basic simulations: period from 12. oct - 13 november. The condition look very similar on the EUR/USD 30 m. period with no time filter.

A simple trading strategy

Cross over of 9 period SMA with 21 period SMA versus 9 Period JMA with 21 period JMA.

Without optimisation the SMA performed slightly better.
After optimization the JMA performed better. (dividing a period of two and optimize on the first part and test on the other part.)

SMA
61, 1 % annual return on account
55,5% % profitable trades
P/L ratio: 2,19
40 #trades

JMA:
47.5 % annual return on account
40,3% profitable trades
P/L ratio: : 1,46
77 #trades

After optimization

SMA
Optimization:
75 % annual return on account
44.8% profitable trades
P/L ratio: : 2,38
29 #trades

Paper trading:
115 % annual return on account
51.7 % profitable trades
P/L ratio: : 2,92
29 #trades

The conditions were
Buy long: Average cross over 2 SMA above 11 SMA
Sell short:Average cross over 4 SMA below 17 SMA

JMA Crossover
Optimization:
110,9 % annual return on account
73.5% profitable trades
P/L ratio: : 2,99
49 #trades

Paper trading:
127 % annual return on account
69.6 % profitable trades
P/L ratio: : 1,82
56 #trades

The conditions were
Buy long: JMA cross over: JMA period 10,6 phase 16 above JMA period 7,6 Phase -38

Sell short:: JMA cross over: JMA period 9,5 phase 0 below JMA period 3,9 Phase -95


The idea is because neither the SMA or the JMA are crystall balls you cannot know which will perform better.
In the example the digital filter performed better but it has a lower P/L ratio. And this was during a period with ranging patterns and a lot of volatility.

So my opinion is that this simulation does not prove anything. But you have to be aware that complex digital filters are not silver bullets and can be outperformed by simple moving averages. And just a hint the tests were done in Neuroshell. This was not the jma that under Metatrader is far away from the original. Guys maybe I am wrong you can repeat the tests.

So did you suppress the future leak of SSA completely ??

Look at it Yourself. In fact I am not sure even if it is a SSA algorithm.
Install the both SSA files into the library folder.