Author's - page 6

New comment
 
ivandurak:

1. You supply a fixed window size to the COM input in your case 40 bars. Imho it is not quite right to draw the current bazaar portrait in some way, in general the sliding window size will be variable, with the condition that it is minimally sufficient. 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. have trained 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 better to fill the map evenly from red to purple like a rainbow, instead of concentrating red with its hues in the centre.

1. Yeah, but you can't do that for variable length windows. I mean, you have to pick up similar patterns via a different method, but in that case, you lose the ability to quantize them in space (which is what PPC does).

2 и 4. I agree, and in Forex the patterns distribution will not be evenly distributed in cells, since flat will prevail. But that's not the main point, we can shape vectors in such a way, that it will be more uniform (perfectly uniform only on random synthetic data with uniform distribution IMHO), but it will become a homogeneous space akin to SB, so I don't even know if it should be aspired to.

3. I've studied this: the cells of SCS are not exactly random, there are repeated alternations, that's what I've tried to exploit, but the question is not fully understood, of course. The point is to reduce a theoretically unlimited number of patterns (or vectors) to a limited set (read: BLS cells) and to trace dynamics on this set.

I myself haven't understood one thing about Kohonen: there are some clusters, they are more or less ordered in space, the ACS projects them (clusters) on a two-dimensional surface, and will ACS always project them equally, will the plane constructed by ACS always pass the same way through the feature space under the same initial condition of training? Example:http://www.generation5.org/content/1999/images/kohonenImages.png there is an aeroplane in n-dimensional space, will the ACS always project it this way and not in profile, for example...

 
alexeymosc:

1. Yes, but for variable-length windows, you can't get enough RMS. I mean, you have to pick up similar patterns using a different method, but in that case you lose the ability to quantize them in space (which is what a VLS does).

What's PDA? Please decipher it.

I only found"Investigative Committee under the Prosecutor's Office of the Russian Federation( ICPOof Russia)", so I have a feeling that something is wrong here).

 
her.human:

What is a PPC? Please decipher it.

I only found"Investigative Committee under the Prosecutor's Office of the Russian Federation(ICP of Russia)", so I have a feeling there is something wrong).

Kohonen Self-Organizing Feature Map, aka SOM.

) It's funny about the investigative committee.

 
alexeymosc:

Kohonen Self-Organizing Feature Map (SOM).

) It's funny about the investigation committee.

I see.

SOM distributes the patterns in its own way. How to interpret them is still unclear to me.

Even after we have calculated all of the patterns on the history it is not clear what to do with them afterwards. If the current pattern on the history shows in most cases to buy - buy or sell.

I made an Expert Advisor (in the trailer).

What the Expert Advisor does:

- Memorizes all current patterns, which are made up of 10 different binary signals (you can choose from 17 variants so far),

Altogether we get 2^10=1024 different combinations of signals, buy and sell signals for each pattern are added up separately,

- Old patterns are gradually forgotten as new ones arrive (forgetting is regulated in the settings),

- We calculate the ratio of signals for each pattern, the type of which is outweighed (buy or sell), the signal is formed in the range from -1 to +1,

- Then we make a decision to enter, exit or reverse,

(here I do not know how to do it better, may be you can advise me how to do it better),

In general counts patterns in a direct way without GA and generalizations COM.

You can add variants of signals, the number of signals at the input (to increase the size of the input vector), or even input the outputs from COM.

Who is not lazy to try, may have thoughts on improvement.

I will not draw beautiful pictures, try it yourself).

Files:
Indic_Stat.mq5  11 kb
 

ivandurak:

2. If you squeeze the chart to the limit, three areas of flat, trend up, trend down will clearly stand out on the history. I will not try to formalise them, 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.

IMHO, it should be like this - flat/trend/random wandering. For Dow Jones shares the percentage is about 35/40/25%. so before looking for patterns and other things, you need to identify the current state of the market.
 
Dima_S:
IMHO, it should be flat/trend/random wandering. For Dow Jones shares the percentage ratio is about 35/40/25%. so before looking for patterns and other things, you need to identify the current state of the market.
Could you elaborate on how you managed to identify 35/40/25%? And what can it give you for trading in the future?
 
her.human:

- Old patterns are gradually forgotten as new ones arrive (forgetting is adjusted in the settings),

Eh, can't run it in the tester, but from the code I like the idea very much, as a simple implementation of adaptive approach and learning.

Separately, I would like to mention forgetting. In this connection I think it is correct to perform forgetting by all patterns on each new bar, not only by the current one:

// Было
// arr_buy[index]*=forgetting;   // забываем старые паттерны, если надо
// arr_sell[index]*=forgetting;  // забываем старые паттерны, если надо

// Стало
for (int i = 0; i < ArraySize(arr_buy); i++)
{
  arr_buy[i]*=forgetting;   // забываем старые паттерны, если надо
  arr_sell[i]*=forgetting;  // забываем старые паттерны, если надо
}
 
hrenfx:

Eh, can't run it in the tester, but I liked the idea in the code as a simple implementation of adaptive and learning approach.

Separately, I would like to mention forgetting. In this connection I think it is correct to perform forgetting on each new bar by all patterns, not only by the current one:

Why not? It runs normally and is even optimized. You can run on opening prices.

Forgettable: I don't think it will be correct on every bar, some patterns are very rare and can be forgotten completely.

But you can experiment and maybe I'm wrong.

PS. remembered that your terminal is not installed.

 

If they are very rare, they are rather the exception to the rule. So they will be filtered out at once.

P.S. It is advisable to put a time filter

Более того, рыночные закономерности колоссально зависят от времени суток, сезонности и т.д. Поэтому при их поиске следует отдельно учитывать временные зоны. Прибыльно-используемая крайние несколько лет ночная торговля некоторых кроссов - яркий пример наличия РЕАЛЬНОЙ закономерности, которая присутствует лишь только в определенном интервале суток. И ее никогда бы не нашли, если бы исследовали весь исходный ВР, без фильтра временных зон.

and try different characters.

 
hrenfx:

If they are very rare, they are rather the exception to the rule. So they will be filtered out at once.

P.S. It is advisable to put a time filter

and try different symbols.

The time filter is set - these are signals from 15 to 17, the Expert Advisor analyzes them too, maybe a little crooked, I have not perfected it yet.

It is not difficult to add different symbols, I initially wanted to do so, then I commented it out. Maybe I will add more.

I have more of a problem interpreting the obtained statistics.

The output contains a signal from +1 (all the patterns worked long) to -1 (all the patterns worked short).

For example, the statistics (signal) is +0.7 (some people call it a probability), i.e. most of the patterns worked long.

What should I do? Should I buy, sell, wait for a stronger signal or get out of the market?

12345678910
New comment