Neuromongers, don't pass by :) need advice - 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
OK, let's forget about the training one for now.
I have a network approximating 1 value, which is whether price will go up or down. I will not go into details.
We can take a test sample (not those data on which the network learns but the data where the training of the neural network stops when the error on this sample stops falling) as a single piece and this piece will come right after the training sample and before the OOS period where the network actually trades. In this case we estimate the network generalization error on data that may differ significantly from the training array.
Another option I mentioned is to generate a control sample randomly from the array of samples going before the OOS. The samples would be shuffled. ) So it turns out (at least in my case) that the adjacent samples are similar and the network learns on one sample (training sample) and immediately the network generalization error is estimated on the adjacent (control) sample. In this case, the error minimum on the test sample can be much deeper than in the case when the test sample is taken by one separate piece.
It's more or less clear now. I have no training. At all.
So there can be no test sampling in the form you suggest.
To be honest, I doubt it would be very effective at all.
I am not against its application, but I don't see an effective scheme.
To be truly testable, the sample must be not on the training interval. I.e., well before the OOS. Which essentially gives a delay in use and no guarantee of improvement :) . Test sampling is good when you train once, test it and then use it dumbly, I don't see the point of using it to predict price series.
Don't take it as a reason to get rid of it (this is an appeal to all readers :) ), it's just an imho.
It's all reasonable what you said. I just want to test my assumptions with practice, although I myself also go from one option to another and then come back... ))) But have made it a rule to always use a test sample in some way.
I also, by the way, always check the system on a forward test before placing an EA on an account. I can tell you that all my EAs pass a single forward test (I've had enough time to do multiple forwards only for EURUSD H1). If they didn't pass, I don't even bother putting them through, because I no longer have any confidence ))))
Another spoonful of tar. An individual net.
You can see that the average is +, but, holy shit, the spread is not small at all.
Another spoonful of tar. Separate network.
Of course you can see that the average is +, but, holy crap, the spread is not small at all.
Hello!
Have you done any devolatilisation (found in the articles) for the inductors you feed to the net input?
You could also try to make the inductions devolatile.
Alas... I don't serve any indices.
Interesting...
What about normalisation to a single interval (or at least an interval)?
Dare I ask an immodest question: what's the teacher? maximising profits through forecasting a few bars ahead?
Dare I ask an immodest question: what's the teacher? maximising profits through forecasting several bars ahead?
No, this approach does not integrate well with the echo network. Already said in this thread about the teacher.
What about rationing to a single interval (or at least an interval)?
Well, if it's just like that, yes, I do :)
We forgot about our "sheep"). The starting message is "How to improve?"
I suggest we abstract away a little (just a little) and think about how to improve the NS result in general, and not just this one, in the application area we are interested in. Here, point by point:
1) Choice of inputs/outputs (a matter of intimate and almost always not subject to discussion, in this case, it is based on a theory approved by two experienced in our business forum members, and we believe that there is nothing to improve)
2) Inputs preprocessing (the question seems rather simple and quite open, we can discuss if it will be known, what and how is done in this case (although I have a sensible NS which basic "zest" is original (not met anywhere) coding of the input data, which I am not going to share yet))
3) Mathematics of NS. (Everything was invented here before us. You may feel free to share and discuss anything you like. Except that all attempts to improve something here are more like shamanism and blind experiments than conscious action)
4) "Organizational" questions of NS. (How / when to train / retrain, periods / intervals, the logic of the network output interpreter, MM, etc. We have seen reports on the entire TS in general. Saw, but sensible ideas for improvement by looking at reports other than some trivial MM comes to mind.
What have I missed?
Where/what theoretically can be improved by taking advice of a person who is not immersed in the essence of development? That leaves items 2 and 3. Point 2 is omitted by TopikStarter as not worthy of attention "there all the usual" (although in my opinion, there may be variants). Point 3, there are articles which you can not understand without 100 grams, and which I personally can not yet fully comprehend (an attempt to implement even a simple echo network has failed so far).
TheXpert, can you tell me something else about your TS that is not a secret? We would a priori be interested, as you have a certain result (I personally think so), and it may "backfire" with smart advice. For example, I wonder:
- Why the "echo"? You've been there, tell me about its pros and cons. How did you dig it out in the first place?
- Inputs/outputs: Mr. joo talks about "flowing patterns" and type 2 TC. I think "flowing" is worthy of discussion, the 2nd type is just an evil one imho.
a) Are you really sure that the inputs/outputs cannot be improved?
b) Pre-processing: How does it look like? Has there been an analysis of the distribution of the input values, for example?