Discussion of article "Practical Use of Kohonen Neural Networks in Algorithmic Trading. Part II. Optimizing and forecasting" - page 3
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
In the paper, during the analysis of different maps, the property that in two maps a cell with matching coordinates (X; Y) corresponds to the same set of TC input parameters is used. How is this rule formed?
If maps are compared, each map is a slice by the i-th weight coefficients of the corresponding neurons (the same pair of X;Y coordinates is the same neuron). And each neuron collects similar (but not necessarily exactly the same) vectors, i.e. in our case TC settings. This generalisation is guaranteed by the Kohonen algorithm itself.
I've been interested in SOM for a long time, I found time and read the primary source book "Toivo Kohonen: Self-Organising Maps".
The article is certainly informative, but for some reason I think the author is wishful thinking.
That's what Kohonen says:
i.e. all that SOM can produce is visualisation based on input features, not prediction.
I've had an interest in SOM for a long time, I found time and am reading the primary source book "Toivo Kohonen: Self-Organising Maps"
The article is certainly informative, but for some reason I think the author is wishful thinking.
That's what Kohonen writes:
i.e. all that SOM can produce is just a visualisation of the input features, not a prediction.
Kohonen maps allow forecasting by analogy with the reconstruction of incomplete data, i.e. by generalisation. This is a reality, available in large volume, in particular on google under the query "kohonen map time series forecasting". The above quote confirms rather than refutes the author, and looks like an attempt to juggle terms out of context. In the article we summarise and cluster the data, which makes it easier for us to make decisions and interpret the market (instrument) in preferred states (classes). There may be a positive effect of moving from SOM to LVQ, but I don't have such material. However, as learning without a teacher, SOM is in some ways more suited to the "black box" of the market, imho.
The above quote confirms rather than refutes the author, and smacks of an attempt to juggle terms out of context.
It's up to you how to take my statement
I like the idea of using SOM as a decoder(electronics) very much: when training, we feed a set of features (meaningful based on our vision), the output is a visual evaluation of it.... then using the trained SOM in trading,
but SOM does not work as a classifier, after training it will not give the same picture on new data as on the test.
In the article we generalise and cluster data, which simplifies our decision making process and interpretation of the market (instrument) in preferred states (classes).
Yes, this is exactly what SOM does - it can be used simply to find and analyse patterns in the original data.
but this task also has no solution "head-on", for SOM training we submit data based on our unsupported hypothesis about the existence of regularities and without knowledge about the amount of information in each data set, if one of the data sets contains excessive information, it will lead to distortion of the result, as the simplest example - we use some signs on TF H1 for a set of trading instruments, but we did not take into account that the time of trading sessions for some instruments is 1/3 of a day, and for others it is 24 hours a day.
--------------------------------------
there are a lot of medical studies on the Internet with SOM application, for these purposes it is a correct application of SOM - data sets and features on which medics want to cluster and visualise their data they prepare themselves based on their knowledge.
As applied to trading, the use of SOM from the medical point of view will look like this:
- we have statistics of patients with snot, appendicitis and limb fractures.
- We apply SOM and conclude that if a patient comes in with a runny nose, the patient is likely to suffer a limb fracture and appendicitis is statistically unlikely.
- we conclude that a patient with a runny nose is more likely to sneeze in winter, which leads to falls and causes limb fractures.
Well, this is all logical from a statistical point of view, and even logical, but to my mind to conclude that a patient with a runny nose will lead to limb fracture in the future and to take preventive measures based on this study is not the best way to use SOM.
The analogies are drawn by ear, imho. Besides, in the best traditions of trolling, the criticism is based on assumptions not taken from the article. It would be desirable to write not in a nihilistic style (like you can't do this (what exactly - still not clear)), but constructively - how and what should be done (NB, we don't have "supported hypotheses" in advance - if we had them, the network could be not used or you could use the network with a teacher).
criticism is based on premises not taken from the article.
Your article is based on your search for "a dark cat in a dark room, especially if it is not there".
The analogies are drawn by ear, imho. Besides, in the best traditions of trolling.
I don't see the point in continuing the discussion in this direction, there is Kohonen's book, there are good introductory articles on SOM on other resources (for some reason I trust basegroup.ru - interesting materials) and there are no forecasting tasks for SOM and there is your article, which like Wiki claims that forecasting with SOM is possible.
I don't see the point in continuing the discussion in this direction, there is Kohonen's book, there are good introductory articles on SOM on other resources (for some reason I trust basegroup.ru - interesting materials) and there are no forecasting tasks for SOM and there is your article, which, like Wiki, claims that forecasting with SOM is possible.
I have nothing against basegroup, I've known them for a long time, but the statement "there are no forecasting tasks for SOM" is not true.
I have nothing against basegroup, I've known them for a long time, but the statement "there are no forecasting tasks for SOM" is not true.
use search, in google it will be the query string "som forecasting site:basegroup.ru".
the search engine will give a description of SOM, which does not have forecasting tasks.
the search engine will give you user questions about using SOM, but the answers will also justify that SOM performs other tasks.
SOM can simply visualise multivariate data, SOM map analysis is a heuristic task rather than a formalised one.
The screenshot above, which I posted when I started the discussion with you from the book "Toivo Kohonen: Self-Organising Maps", also suggests only visualisation tasks for SOM.
but Wiki and your article suggest using SOM for prediction tasks, it remains to find the truth.
Your judgement and inability to use a search engine is clear. I have voiced all my arguments. No need to litter the discussion. I will answer in your own words:
I see no point in continuing the discussion in this direction
In your own words:
similarly, your responses to my posts are:
to your attempt to juggle terms.
Also, in the best tradition of trolling.
ZY: did I give you a search string for google? what search are we talking about? about the quality of material on basegroup.ru you and I have come to a common opinion, or do you suggest to search on Wiki, the same Habr, quite often good discussions on Stack Overflow .... but they are not always reliable sources