Kohonen and Patterns - page 5
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
ANG3110 if it's not a secret how did you filter the data smoothed or what?
Yes, I smoothed it out a bit. I use several filtering methods. In the case shown in the 3-month forecast chart, the regressive adaptive filter was used.
This is what it looks like without any forecast:
I used a General Regression Network (GRNN) to build the above pictures. This is a modification of the probabilistic PNN designed for approximations and predictions. What's a variant of a fairly accurate estimate of the network prediction right off the bat in prediction? Describe something, you're already on page 2 hinting at something, but haven't written a single word on the merits. I have worked with almost all major networks and I think to understand what I am talking about is not difficult.
I hope to continue the conversation when your busyness will be lessened.
In the case shown in the 3-month forecast graph, the regressive adaptive filter was applied.
The filtering looks very good. Can I have more information about filtering algorithm?
The filtering looks very good. Can I get more details on the filtering algorithm?
Take 2 arrays a[] and b[]. Let's put Close[i] into them. Then a short period of linear regression N is taken and raced in one direction. At every following step the data is summed up. Then it is raced in the opposite direction. The same thing is done. And so on (sort of smoothing out). Then forward and reverse runs are summed up and averaged (a[i]+b[i])/2. aa and bb are linear regression coefficients.
It turns out that this muving is overdrawn. Is that why it looks good?
It turns out that this muving is overdrawn. Is that why it looks good?
Well, it's not a muving as an end point of indicator type - it's a filter. In addition, the network needs really centered data, and any non redrawing indicator - muvinga type or any other indicator is not in the middle of the data, but offset, and besides the end is always wobbling. For the network, this is low-quality data, and it will give the same unrealistic picture on such data. A normal moving average is half a period off centre. EMA by a third and LWMA by a quarter. The linear regression is shifted at an angle, and its horizontal shift is variable, but it is still shifted. But such a filter is well centered exactly on the current data that is fed to the network input. But by the way, if it is not redrawn and used as an endpoint, it will be close to the ordinary linear regression, only smoother. The network is redrawn pre-flash. If you are interested in non redrawing indicator with very good smoothness, it is T3, but it is lagging a little, it depends on b coefficient. But fast indicator with less smoothness is DCT.
I understand everything.
There are a few minor disagreements on my part, like:
...неперерисовывающийся индикатор - типа мувинга или любого другого проходит-то не посередине данных, а со сдвигом, и к тому же конец всегда болтается.
It's not supposed to hang around here.
The normal muwings, on the other hand, are half a period off centre. EMA by a third, LWMA by a quarter.
EMA and LWMA are a recursive type of digital filters. For this type, in principle, you can't define the notion of a smoothing window width, so it's not appropriate to talk about "centre" and "period". You can talk about group delay and phase delay.
Although, that's just for my own relevance :-)
ANG3110, in a neighboring thread I was discussing about alternative representation of algorithm's predictive properties, may be we can use your NS for that? Very curious results you get.
Thanks for the invitation, maybe I'll check it out at my leisure. It's the same with writing, you write something once and then you're drawn to it. That's good if it's a break from work. But it can be distracting, at least for me.
About shifts of different types of muwings...
Here is the script that I attached and you can see what shifts are really occurring.
If for example it is Daily and period is 5, then the end will visibly wobble on the 0th bar in indicator mode.
Good afternoon.
Has anyone tried to search for patterns (on MACD for example) using Kohonen's net?
If so, please share your impressions and experiences.
If anyone has similar ideas - I invite to chat. Preferably here, if serious, to email.
Please write about the theme and to be specific.
The result of New Year's rest:
A little later I will post in the base.
Who wants to join the research? Please express your thoughts here.
I will ignore posts like "may I add stochastic here" or "make fixed stops with trawl in . points" I will ignore.
I appreciate constructive suggestions and criticism.
I want to provide you with some details about the Expert Advisor. This is an ordinary Expert Advisor by signals. The AutoMACD indicator gives the signals.
About the indicator - this is a simple variant of the adaptive signal indicator.
Signals are generated according to pattern statistics.
Patterns are clusters of Kohonen nets.
The Kohonen net (patterns) is trained (reformed) parallel to price formation and quite fast.
It will become clearer when you look at the code. I apologize at once for a little complicated logic, this is a working draft.
if (Volume[0] == 1) what is this condition for?
I understand that all the processing is done in the indicator? very interesting, not quite figured it out yet, thanks