NS + indicators. Experiment. - page 8
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
NSDT does not have Kohonen, but there are other classifiers in the Adaptive Net Indicators addon. Although you can connect NeuroShell2 net directly to MT4 or NSDT.
I wonder how do you plan to use Kohonen's net for trading? It's not three classes there, it's much more. I need to think of some algorithm to parse the obtained classes.
When I first started research on bar classification using Kohonen Maps, I coded bars as follows to simplify the experiment: Bullish +1, bearish -1, dodge 0. I arbitrarily created 15 classes in NS2. After training, I ended up with 4 classes empty. I revised the number of classes to 11 and filled this grid in МТ4. Then, I used the script to enter three classes near all extrema and performed a simple reverse operation. The input data was "as is", i.e. just class numbers. There was no normalization at all. Then I "set" a simple Expert Advisor consisting of those grids only, using signals from the reverse direction for 2005. Strangely enough, it even worked. Of course, I didn't even try to make any conclusions about the results, since it was not even an experiment, but a laboratory work. But I did want to think about the prospects for this project.
I mean, I didn't need any class selection algorithm. Everything was clearly drawn on the graph in NS2 itself. And, in general, it is possible to achieve the desired result in NS2 itself. But to do it, you have to jump back and forth a bit, play with parameters, look at graphs, look at output data. So, it's a bit tedious, but quite possible.
I wonder how you plan to use the Kohonen network for trading? There are not three classes, but much more. You must think of some algorithm to parse the obtained classes.
Certainly not 3 or even 10. At the first stage I analyze them manually to come up with appropriate criteria. Then I automate it. Of what is planned:
- each input is evaluated by "degree of idealism" - how close it is to local extrema (33) for example
- then each class may be evaluated from the viewpoint of purity - how many inputs are ideal or close to it and how many are empty.
The first study of the first version of the set of inputs (neurons) showed that there are 2-3 classes that are quite acceptable from the point of view of market entry. And the power of these classes yields about 2-4 trades per day.
Everything is still raw and under development :-) Then I plan to make (already started) a program which will independently create a network, train, estimate, run in the built-in tester, save the results, create a new one, run, select the "champion", etc.
ZS. Ok, I'll torture the NS2 for now, only it is miserable, as I said, in terms of interface and evaluation of network training results.
Why not? There is.
When working with a Kohonen network, we cannot tell in advance how many classes we will get. In the learning process, the input feature vector will be divided into classes, as long as the given separability criterion is satisfied (e.g. Euclidean distance). Suppose we got "clusters", find centres, and then...
Visually, by map, of course, you can estimate, but it would be desirable to automate. The second step should be a teacher who would indicate which classes, which actions correspond to what.
I am still raw, too. There are ideas, developments. I am also working with NS2.
When working with a Kohonen network, we cannot tell in advance how many classes we will get. In the learning process, the input feature vector will be divided into classes, as long as the given separability criterion is satisfied (e.g. Euclidean distance). Suppose we got "clusters", find centres, and then...
Visually, on the map, of course, you can estimate, but it would be desirable to automate. At the second stage, there should be a teacher who would indicate which classes, which actions correspond to what.
I am still raw, too. There are ideas, developments. I am also working with NS2.
Well, why can't we say how many classes we'll get? We can tell you everything. I don't know about NSH2 (I can't say for sure - I haven't seen it), but in Trader you can specify as many classes as you want - buy class, sell class, quit buy class, quit sell class. You will have as many classes as you need. And there is no problem with it.....
Firstly, I don't like NS2 because you can't do Kohonen map colouring there, and secondly, you can't change the clustering on the fly, which the deductor easily does. You don't have to retrain the network for that! Just change the parameters for combining into classes. In the same deductor you can set the number of classes, you can set significance levels for clustering, then it is really unclear how many of them there will be, and you can do without classes at all, and look at what cell the input falls into.
www.basegroup.ru is also easily searchable by google
www.basegroup.ru is also easily searchable by google
www.basegroup.ru is also easily searchable by google