Optimized Trend Trading

Has anyone ever tried a machine learning approach to find a model the gives +/- the same shape as the non-causal smoothers (Hodrick-Prescott, SSA, ...). The inputs of the model would be a very good filter like the JMA, and also technical analysis indicators, or other variables that could possibly be related to the non-causal smoother, and the output would be a very smooth, low lag filter.

btw: before anyone mentions neural nets again: there's much better techniques these days!

I work with R; If you're interested, send me a pm

Hi it is not obvious to make the digital filters work. If you use a digital filter and apply a simple Cross-over strategy it is a waste. The simple SMA cross-over is giving more reliable results recently (I made an experiment SMA crossover expert versus JMA cross-over expert. The JMA was a lame, the SMA gave much more robust solutions
http://www.forex-tsd.com/manual-trad...ystem-120.html

I published there my work on the digital upgrade of ASCtrend and unfortunately I had to prove that the mere fact we use a JMA does not mean we are making any better, however the fractal mod of FRASMA outperformed the SMA original at any condition).

That does not mean that DF are useless, but they need much more sophisticated or different approach.

Here I give an example of someone who did a very interesting trading system using Neuroshell. The strategy is based on a complex mix of digital filters and GA optimization.

http://www.investireoggi.it/forum/ne...i-vt32426.html

Unfortunately a lot of the links are not working anymore.
You need google translate.

Look the rules are as follows:

AMTRADING RULESRules
BUY LONG CONDITIONS: A>B(Lead(JMANSDT_001(Close,9,0),2),JMANSDT_002(Clo se,19,0)) Not(A[b]=B>=C(0.00539,Lead(JMANSDT_003(Add2(Sub(LinTimeReg PredValue(Close,55,30),Hodrick-Prescott Window(Close,100,500000,0)),Mul2(1.715,Sub(Hodrick-Prescott Window(Close,100,500000,0),Ehlers Trend Line(Close,3,3,3,3,60,40)))),5,0),10),Neg(0.005... ...39)),A>B(Lead(JMANSDT_003(Add2(Sub(LinTimeReg PredValue(Close,55,30),Hodrick-Prescott Window(Close,100,500000,0)),Mul2(1.715,Sub(Hodrick-Prescott Window(Close,100,500000,0),Ehlers Trend Line(Close,3,3,3,3,60,40)))),5,0),10),Add2(Sub(Lin TimeReg PredValue(Cl... ...ose,55,30),Hodrick-Prescott Window(Close,100,500000,0)),Mul2(1.715,Sub(Hodrick-Prescott Window(Close,100,500000,0),Ehlers Trend Line(Close,3,3,3,3,60,40))))),1) A>B(Momentum(JMANSDT_004(Close,3,0),1),0)

SELL SHORT CONDITIONS: A[b]B(Lead(JMANSDT_003(Add2(Sub(LinTimeReg PredValue(Close,55,30),Hodrick-Prescott Window(Close,100,500000,0)),Mul2(1.715,Sub(Hodrick-Prescott Window(Close,100,500000,0),Ehlers Trend Line(Close,3,3,3,3,60,40)))),5,0),10),0.00539)) IfThenElse(A>=B>=C(0.00539,Lead(JMANSDT_003(Add2(S ub(LinTimeReg PredValue(Close,55,30),Hodrick-Prescott Window(Close,100,500000,0)),Mul2(1.715,Sub(Hodrick-Prescott Window(Close,100,500000,0),Ehlers Trend Line(Close,3,3,3,3,60,40)))),5,0),10),Neg(0.005... ...39)),A<B(Lead(JMANSDT_003(Add2(Sub(LinTimeReg PredValue(Close,55,30),Hodrick-Prescott Window(Close,100,500000,0)),Mul2(1.715,Sub(Hodrick-Prescott Window(Close,100,500000,0),Ehlers Trend Line(Close,3,3,3,3,60,40)))),5,0),10),Add2(Sub(Lin TimeReg PredValue(Cl... ...ose,55,30),Hodrick-Prescott Window(Close,100,500000,0)),Mul2(1.715,Sub(Hodrick-Prescott Window(Close,100,500000,0),Ehlers Trend Line(Close,3,3,3,3,60,40))))),1) A<B(Momentum(JMANSDT_004(Close,3,0),1),0)

After all those conditions a Genetic optimizer is performed.

http://www.investireoggi.it/forum/ne...t32426-16.html

Here it is the template for Neuroshell.

This is one of the most successful strategies for digital filters I know.
The Neuroshell is just a tool you can find some useful ideas of what inputs were used, the optimization is different drink. I like very much this program but I feel uneasy to rely on a proprietary software.

R is a cool program I have installed on my PC but it it requires a lot of effort, the learning curve is steep like a mountain for a guy with no programming skills whatsoever.

But you can do a lot of fancy things with it. Recently I worked on Brain Trend with digital filter, the code is open source, look at it, it is a sophisticated approach of using digital filters. It gives interesting results even without optimization.

I have a particular opinion about optimization. What we do is basically a statistical optimization. Whatever algorithm, you use you always optimize the Balance, Profit to Loose Ratio, Maximum return on account.

And THAT IS WRONG. THAT IS WRONG because the markets are not GAUSSIAN. The markets are soooooooo complex that sometimes even the most powerful computer on the planet will do no better than an Elliot Count. Sometimes all the algorythms will find the solution and we will have a low phase space singularity.

For example an Eliott count guy sees that the patterns of the impulse is over. He sees strange range movements and says. Aha correction pattern. Is he able to predict very well, I doubt but he can very well adapt its trading strategy to the market conditions.

The guy with the optimizer has just finished a perfect optimization and starts to trade an optimized solution for a trend in a ranging market. What happens next?

Are the algorithms able to make face recognition as good as humans? And appreciating the market state requires the same cognitive process as the face recognition.

Here I do not mean that algorithmic optimization is useless, I mean that its mechanical use leads in dark places, whatever algorithm you use.
The human and the machine have to work together we have no better answer yet at our retail level.

Hi John Last

can you explain the concept behind your JMA Based BrainTrend Signal and stop line. I like to understand the reasoning behind how things work. Thank you.

Hello!!

There aren´t people like john last.

Great work. Really thanks.


But, why all the people want to filter the price ??

Sorry but there aren´t perfect filters without repainting or lag

Best regards.

Imagine a football game. But strange football game because you cannot see the players you see just the ball.

You do not know when the ball will be kicked, or its direction. You do not see the players of the either team.

The filter would answer the question. Is the ball kicked by another player, or it was going up and down as a result of previous kick?

That is is the logic of brain trend. We measure the ATR and we want to know if the last kick is by a player or just a noise.
In brain trend we have three inputs: Stochastic, ATR and Digital filter JJMA. Basically it is a variation of the volatility system of Wilder (he considers the volatility system its best).

I will try to make 101 on noise. The moist important distinction is between the black noise and the pink noise. They are both dangerous the black noise can be like a shark, the pink noise is like many piranhas.

...and R can easily be integrated in MQL4 programs:

looking forward for your 101 lesson. great works!

Hi,

I will share with you my impudent Newby engineering mod of jjma.
So there are two mods. One is close to the jjma in smoothing but reacts better when it have to. Both use nonlinear adaptation to the price.

The other is really different. But unfortunately I cannot use it to full power because the calculations are done within MT4 and they are very intense for MT4 and it crashes when more power is applied.

Those movings are pretty useless on their own. But when I insert them into Brain Trend I can add more smoothing and my response is quicker when the market dictates so.

I did not forget the 101 but I do not have time right now.

John,

any reason why you use period "25" for the jurik ? My charts runs with period "5" (1h chart)

Sixer

The reason is to show the difference. Because if the period is small the differences will not appear as clear.

The idea is to use more smoothing but to have a sharp reaction only when you need it.

Anyway I watched there is a difference too.

Sometimes you do not see the blue line of JJMA because it is covered by the red.

I will disclose the idea. The idea is that the digital filters are used on their own and the fractal moving averages are used on their own.
So I asked my self the question what would happen if we combine those advanced approaches.
The sole problem is that I am not a coder, not engineer or mathematician.

I really wanted to know badly what would happen. So I made it simple I used as a carrier a fractal moving average and I replaced the iMA function inside with iCustom function of JJMA. So you now have the results. The two moving averages are different because they use different formulas for the estimation of the fractal dimension. The yellow line use a R/S analysis, the Red line uses the normal FGDI formula.
So when the market is calm it is a normal filter. When we have a lower dimension with increased probability of black noise the filter tends to react quicker.

So when we have a lower fractal dimension the movement is persistent. There is a bigger probability that the next movement will be in the direction of the previous. So when there is a reversal, the reversal tends to be quick and abrupt (Black noise). So that explains how the V tops and V bottoms are formed.

On the other hand when the fractal dimension is higher than 1.5 we have a bigger probability that the next movement will be in the opposite direction.
And the price goes up and down in a lot of oscillations. There we can find pink noise, a lot of whipsaws up and down.


If the price time series were random there would be no correlation of each movement with the previous movement.

The brain trend still amazes my how it is able to find a solution in very hostile environments. Normally it should be used when there is a low dimension in both two key levels 15 and 30.

I can give a lot of examples of this. This approach is different and independent from the TA perspective so it combines really well with it.

H - Hurst exponent
When H = 0,5 the FGDI = 1,5


Pink noise 0<H>0.5
Black noise 0,5<H<1.0

See chapter 13: Fractional noise ans R/S Analysis
From the book fractal market analysis.

In this book only the theory is discussed without any practical consideration whatsoever. Nevertheless this knowledge have very practical use in the TA.

i dont think u need an ea to trade the moving average,

This thread has been designed as a general discussion of optimized trend trading.

My hypothesis is that the digital filters need more sophisticated approach, a simple cross over does not seem to work.

I have several points:

1. However the example of Brain Trend showed that a digital filter can substantially enhance the capability of a more complex system.

2. The digital filters can be enhanced with a combination of adaptation with fractal geometry. The shots some posts ago were based on that approach. I will show we that it was insanely simple to do. In fact it is done you need to combine existing things.

And now I will show another example how a digital filter can enhance an existing strategy.

This concerns the Trend magic indicator. I just replaced the normal CCI with something else. And we have nice results. You can adapt the filter and the period. In the original it was set to 50 period CCI as a proxy for trend measure.

As usual you need to install the JJMAseries in the include folder.

As a joke I call it Trend Madic. But it is a digital edition of Trend magic.
You can play with:
-length of the CCI
-JJMA smoothing
-phase of the JJMA smoothing

So you can more or less reactive at will.

Enjoy

I forgot you may need the jjma to run it.

http://articles.mql4.com/310

This article is very important. I have to repeat you need the jmma series include file to run the beast.

Is jjima an indicator then?

hi, John, thanks for all the information here. one question about your indicator (#_Fullssa_Normalize), are you the creator of this indicator? or you got from someother site? if it is latter , could you point me that site ,thanks. I right like this indicator and I want to learn more about it if it is possible. thank again John.

michaelwen,

check the following links for more information with regard to SSA:

http://www.math.uni-bremen.de/~theod...eople.Research

Sixer

Guys this indicator is in the mql official codebase.
I am not a creator of anything. I just want to combine existing things because there are many unexplored possibilities. Alvin Toffler says that knowledge is usefull when it is combined with another knowledge. Exactly for that reason we can combine our digital filters with more complex systems in Metatrader with just one line of code, but can you combine the original jurik as easy, and with what because you have closed doors everywhere.

Kositsin in the link has created a whole library explaining that every iMA and any averaging algorithm can be replaced by its digital smoothing. The results may be astonishing or may be not. The TSD innovation is to use double smoothing and see what is better. As a resuly I asked my self that we can use a fractal adaptation of jurik smoothing and double smoothing.

How it is done. You have the FRASMA you change the iMA inside of it with the jurik smoothing. You do the same with RS_FRASMA and every other fractal moving average you have at hand.

The SSA normalize, showed that if there is not a reliable way to use SSA without recalculation there is always the possibility to extract the information step by step. The SSA normalize uses arrows, the TSD way is better.

OK Let see an example in real time.
On one shot we have the double smoothed. It is better when the market is more volatile.

That is why you need the SSA squeeze on the bottom in order not to trade against the slope.

The fractal mod of the doubled smoothed is somewhat inferior in reactivity but it is intended to work on all market conditions. And the stops are far too close.

In those toys we have:
Digital filters
Fractal geometry
SSA
Use of the volatility

This is fairly complex toy. What I like it that it adapts to the market and do not try to predict.
What I would like to add is a statistical volatility changing threshold according to the hour. Because the brain trend responds primarily to volatility. And the volatility changes on every pair in a different way during the time. So we can add some statistical knowledge into it.

So it is better to optimize those toys for time filter. It is unwise to look for tops and bottoms it is better to see how it performs in a specific time for example the European session and compare European session with European session. That is easier.

The second thing let not forget the fractal dimension characteristics. If we have red on both 15 and 30 time frame our Persistent Movement Roller Coaster can begin. We will show to the old guys how we can pick tops and bottoms. If we have on both 15 and 30 time frame high fractal dimension that means that the phase space is deemed to be too complex. I call it high phase space singularity. The phase space is so complex that nobody is right. The price goes high and low like crazy. Do not expect Brain trend to perform under those conditions, nobody can. In those conditions statistical methods are used, because the phase space space is so complex that we can analyse it as Gaussian.

Have a nice day. I will show posts of high phase space singularity and low phase space singularity. This is something complete new.
But try yourself use the FGDI and look around.

Hi,

I want to post another example of Low phase space singularity. What is typical of that is that many different algorithms are able to find the same solution simultaneously and act accordingly. Look at the shot is it an accident that the Brain trend, the ASCTtrend stops with digital smoothing, the ASCtrend signal, the Trend magic with or without digital smoothing and the SSAsqueeze find simultaneously the solution?

You see brain trend had good results, ASCTtrend had good results, Trend Magic had good results. A simple moving average will have good results. The human trader can go against the trend and can be hurt. Price action specialists will see a break of a support (where the Brain Trend reacts and that is true), and that is why they tell trade what you see and not you think.

The idea is that when we have a relatively simple phase space of the possible solutions the algorithms are able to find simultaneously a solution. The algorithms are able to cooperate, but the humans we do not cooperate (one guy think it is oversold, the other it is overbought, one guy has long term horizon the other has short term).And FGDI in red at both 15 and 30 m. time frame is a good approximation of the possible Low phase space singularity.

This Low phase space singularity is scary for the public policy makers, because all the market participants start to have the same horizon simultaneously.
They are not able to adjust the economies to the market neither to guide the market as efficient as they would like.

Sometimes the Singularity goes into one direction, sometimes we have a series of ping - pong movements. All that is highly unpredictable several movements into the future, try to train a Neural Net to predict those market conditions and you will see, in fact whatever predicting algorithm you try to implement you will fail predicting the low phase space singularity. You have to react as quick and intelligent as possible, and even a SMA is good enough.

However the SSA squeeze still recalculates. They have to bloc the bar once it is finished, otherwise I do not see the point of using bars.