Elliottware: machine learning extension of Elliott Wave principle
Navigation
Practical limitations in the Neural net extrapolation
Practical limitations in the Neural net extrapolation
 |
| От 22 септември 2011 |
I just would like to point out something very important about the application of the principle of uncertainty the technical analysis and the Elliotware approach in particular. Some may think that if they can use advanced machine learning solutions they can totally not take into account the common technical analysis. I think it is true if the model is base on tic data and high frequency trading there are other rules and principles. However when we go higher the technical analysis becomes important. But what I mean. I mean precisely the support and resustence zones.
Those zones are known as decision zones. That means that there a decision has to be made.
And this is analogy with the principle of uncertainly (a quote from wikipedia).
In quantum mechanics, the Heisenberg uncertainty principle states a fundamental limit on the accuracy with which certain pairs of physical properties of a particle, such as position and momentum, cannot be simultaneously known. In other words, the more precisely one property is measured, the less precisely the other can be controlled, determined, or known.
Well this is just an analogy. In technical analysis when we are closer to such a decision zone. The scenarios are clearly measured and predetermined. The price is either doing this or that. But if we know precusely the scenarios we do not know the actual direction.
And vice-versa when we are away from those decision zones. We may know very well the general direction of the market but we have not a clue of the trajectories.
And in that particular situation the Neural Net by their extrapolating abilities are giving us a help. When the direction is clear but the scenarios are unclear.
If you are clearly near a decision zone the performance of the neural net will depend more on luck than other things.
Here I woul like to post my shot. Here we have two scenarios. Those scenarios are clearly cut. One is going high, the other is goind down to the channel. From the downside of the channel a new decision has to be made.
I used BPNN Caterpillar with the same training period on 30 m time frame. Both of the BPNN was with lag bars 40. However one had 3 computations the other 5. One of the nets was pointing high the other low.
I mean a slight change of the inputs and we have a totally different prediction. I think if we consider a decision zone it is normal. Basically what inputs will you use and the outcome is based on luck.
Comments
I would like to add a shot from the Oanda open orders. We have an accumulation of retail orders only at the bottom of the channel. That means that there would not be a stop hunting from the market makers at the up side. So that would mean if we have a break - iyt to the upside that would mean it is something else.
So ... here is the shot.
I just post it because some very technical guys may think that if the neural net is giving wrong signals the they may think:
poor training if the neural net
need to change the architecture
need to change the learning algorythm
need to change the sampling
need to use another machine learning mechanism.
And maybe the reason is that sometimes there are obvious practical limitations when the extrapolation does not work.
I found PBNN or AR or other forecastors online may give out different forecast after a new bar come out.
So, let PBNN to run 10-20 hours (or keeping it running), then, we may can use the forecast. PBNN need data to train itself for hours for it to give out correct forecast. But, if you use supervised training(previous training, stored), then, forecast can be used straightway.
But, PBNN can be trained by selected price data set, then, locked out weights, no training when using. In this way, PBNN can be trained to forecast top/bottom, strong trend/weak trend, etc. market status. Selection of training price data(patterns) may be a big task, take a lot of time. NN is the best for pattern identification. It may not suitable for continue, real-time forecast to fit data like ma.
Forecast 25 bars or 50 bars or 100 bars movement is better in quiet time, eg before European open. In this time, most of big guys' computers in world are doing the same forecast for next coming London movement. So, you can get the same forecast with big guys if you use similar aglo with them. When market open, price move terribly up/down, then, price data is not steady, and hard for your aglo.
I think we think in the same direction. It is imporartant to know what the machines get as a solution. And this is achieved at best when we do not have a lot of volatility. Bigger volatility means bigger chaoticity. And if the chaoticity is big there is not guaranteed that the solution we get is the same as the solution of the big guys.
But what is PBNN?
PBNN? I got wrong name for the nn. BPNN. sorry.
You can use the visual test on any expert. Just plot the extrapolator and watch it what it is doing. You can pause at any time.
Good afternoon. Asian market is coming to open soon. China, Hong Kong, Singapore are all of big Chinese fx force. Japan and South Korea become smaller fx force. All Asian computers of big guys from Global Banks, Hedge funds, etc... are running to do forecast for BIG Tues.
Monday is good for both long and short players. Swing down, swing up.
Here is my AR_Growth model forecast:
http://www.forexstreet.net/photo/zzztesth1euro-ar-growthfunction-1-tues-25-10-2011?context=user
Here is my volatility forecast which is used for target. Most of people, traders, fx tutors, fx guru, ...., etc play focus on stoploss, physciology, money mangement, ...., etc. But, sad story ? often come out.
24 hour volatility forecast:
North: 1.3945 breakout, then, ===> 1.3970; 1.4035; 1.4060.
South: 1.3910 breakout, then, ===> 1.3890; 1.3855; 1.3795.
-------------------------------
Weekly volatility forecast:
North: 1.3840 breakout, then, ===> 1.3940; 1.4030; 1.4100; 1.4160; 1.4255; 1.4360.
South: 1.3680 breakout, then, ===> 1.3580; 1.3520; 1.3420; 1.3360; 1.3260; 1.3155.
------------------------------
Better forecast, Better performance !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
I would like to post a screen shot at the London open. The fractal dimension is very high for effective prediction from me. It is quite antipersistent. According to the research by Bo Qian and Khaled Rasheed and my ow experience the high fractal dimension is another obstacle for good neural net extrapolation.
I think Jaguar has made some mods trying to extrapolate the iVAR.
Here on the shot is is seenthat when the fractal dimension is high the sygnals from the Brain Trend DE are unreliable.
Here is the link to the article. I just want to post the same shot after some moments. Look how the price action is consistent with the fractal dimension. The price action is accupying the space according to its dimension.
And the last thing. Take a look at this shot. Hee we have the same system with the same setting but lower at the 1m time frame. There the fractal dimension is different and guess what the signals generated by the directionnal system are quite accurate indeed.
The key for me is to find the time frame with the best prediction horizon, the lowest fractal dimension and to use a directionnal signal exactly there. The market is changing all the time if you consider the combinations between all the time frames you may think that the market is never the same during your lifetime. I mean the precise combination between the fractal dimension at all the available time frames.
However there are several states that are pretty recurrent.
Yes I know this is not in the books and tutorials of technical analysis.
Jack1 thanks for the forecast. What setting and what mod do you use?