Author's - 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
My implementation of the Kohonen map, first release . So far it can disassemble the colour palette to see if it works. Codes attached
Total Kohonen Game.
The method is certainly interesting, but it is suitable for classifying stationary objects. The matter is that adequate training and analysis of a 30X30 map requires an array of training vectors of about 50000, further in progression. In such a long interval the regularities (supposing they are present but not stable) are blurred and the map gets homogeneous colour examples at 4. Besides the Kohonen map has turned out to be very sensitive to the type of O[i]/O[i-1] data representation, while the same data (O[i]-O[i-1])/O[i], splits the map into two well-defined areas as it should be. Maybe the whole problem is a crooked hand sharpening, but I'm already offended by mother nature for the neural network between my ears.
I'll try the correlation again, later when the missile is formed, and I need to work on the allegarh, because he is offended
It's been a long time since I've been here.
The respected community is presented with an indicator which builds an optimal portfolio according to the principle of maximum trendiness with minimum with minimum variance in the selected area.
This is a release, the codes are not optimized, so please do not spit too much. It would be better to give me an idea what to work with. The work with indicator is described in comments. Example of its work on fig.
The codes are in the trailer.
I'll try the correlation again...
As much as I could squeeze out of the patterns through correlation, I've posted it here.
I'll quote part of my post from my private message:
It was a pleasure to read your post on the subject of patterns. I would like to add some of my own. In this work I found that whichever piece of data is taken, there are quite a lot of similar Pearson QC (> 0.9) intervals that are far apart from each other. The job description in the second video shows the prediction (outside the vertical lines) for each plot (between the red vertical lines). In the left corner just the number of similar plots and their average and best QC are given.For SB, as written there it should be like this:On SB data (random walk) the forecast must be a horizontal line, with its RMS to diverge with distance.
The esteemed community is presented with an indicator that builds the optimal portfolio by the principle of maximum trendiness with minimum dispersion in the selected area.
I don't have an MT5 for lack of use. However, the code is so perfectly commented that everything is clear.
If you throw out the imperfect multi-bar synchronisation, equity calculation, limitation on coefficients etc. The code is, in its rough form, a complete way of building a synthetic with any condition. Where only this line needs to be corrected:
This criterion is clear - to find the maximum stable trend equity of the synthetic on the building interval. And if we spit on the ambiguity of the criterion calculation and consider the issue in general for any condition of the construction of the synthetic, we need to investigate the topic further:
To calculate this even via GA is a sea of time. Therefore it is practically impossible to investigate the question without an analytical solution, but I would like to.
P.S. If I hook up a cloud, maybe it will be possible to investigate.
Is the pattern theme still going on?
I wrote about my findings here: https://www.mql5.com/ru/forum/133209/page5
- Train Kohonen's ACS on patterns (how you form them is an individual question, but an important one)
- Allocate a number to each cell of the ACS (for me it was coordinates, e.g. 3;5)
- entering a position by activating cell with coordinates x1;y1, and closing a position by activating cell x2;y2. In this case, there can be many input and output cells (their combinations are important)
- The implementation (in my case): trained BCS in dll sends to the Expert Advisor coordinates of the cell activated by the current price pattern, if coordinates show an entry, then we enter, then we also exit, if the activated cell shows a close position. You can genetically look through a lot of variations of entering and exiting at certain cells and record the most profitable ones in your Expert Advisor. I was just writing them down manually after analyzing all genetic runs.
Thus, it's not only the entry pattern that's important, but also the exit pattern. The results are quite different. I have the pictures there.
I suggest that for pattern work (and other relatively complex topics) you post a video like this (in HD):
This way you can visually assess the results of the method. To give some insights and ideas from the outside.
No one can do a better job than the author at concocting a similar video for his method. The main condition is no looking into the future.
This way you can see the dynamics, where each frame can count for minutes (depending on the calculation algorithm).
Is the pattern theme still going on?
I wrote about my findings here: https://www.mql5.com/ru/forum/133209/page5
- Train Kohonen's ACS on patterns (how you form them is an individual question, but an important one)
- Allocate a number to each cell of the ACS (for me it was coordinates, for example, 3;5)
- entering a position by activating cell with coordinates x1;y1, and closing a position by activating cell x2;y2. In this case, there can be many input and output cells (their combinations are important)
- The implementation (in my case): trained BCS in dll sends to the Expert Advisor coordinates of the cell activated by the current price pattern, if coordinates show an entry, then we enter, then we also exit, if the activated cell shows a close position. You can genetically look through a lot of variations of entering and exiting at certain cells and record the most profitable ones in your Expert Advisor. I was just writing them down manually after analyzing all genetic runs.
Thus, it's not only the entry pattern that's important, but also the exit pattern. The results are quite different. I have the pictures there.
1. You feed the COM input a fixed window size in your case 40 bars. Imho it is not quite correct to draw current bazaar portrait in some way, in general the sliding window size will be variable, with the condition that it is minimal enough. In addition, the training vector may include not only the price, and everything from interest rates to indicator readings, including the distribution of current orders, the proximity of support and resistance levels, etc.
2. If we compress the chart to the limit, the history will clearly show three areas of flat, trend up, trend down. I will not try to formalize it, I am not that stupid. The task is to mark out these areas and try to identify them at an early stage of their emergence.
3. trained in COM on history. Dreaming of looking at the trajectory of the current moment on the on line map. If the trajectory is predicted, then a profitable strategy can be picked and run in advance on similar historical areas.
4. It is necessary to build a map to the maximum possible even distribution of clusters. The map of my implementation, see fig. above, shows that the algorithm works almost correctly. There is a classification of input vectors. However, imho it would be more correct to fill the map evenly from red to purple like a rainbow, instead of concentrating red with its shades in the centre.
need to investigate the topic further:
To calculate this even via GA is a sea of time. Therefore it is practically impossible to investigate the question without an analytical solution, but I would like to.
P.S. If you hook up a cloud, maybe it would be possible to investigate.
I disagree. You propose to write an EA with a certain number of input parameters, optimizing which you can try to find patterns.
I, on the other hand, think the right way to go is to first do some deep research and only then write an EA with input parameters based on that research.