Machine learning in trading: theory, models, practice and algo-trading - page 619
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
I don't understand about the volume, isn't 10,000 examples of states enough for training?
It may or may not be enough. It depends on what you're teaching and how you're teaching it.
In my first variants ~10000 was about nothing at all. In the latter, after changing the model of learning for the same NS it was all good.
in this variant it will turn out that the model from the static 100 bars will be neutered, and I believe that it will not lead to the desired search of a possible pattern (
I'm saying that the architecture should be constant, and on each new case, you will move the "attention" window forward. This way the seriality and one nature of a particular dataset will be tracked.
One learning tick: i is the current index. input = [i-100, i], output = [i+1, i+6]. And accordingly i is new every time.
I did not tell you to serve the same bars. =)
I'm saying that the architecture should be constant, and on each new case you will move the "attention" window forward. This way the seriality and one nature of a particular dataset will be tracked.
One learning tick: i is the current index. input = [i-100, i], output = [i+1, i+6]. And accordingly i is new every time.
Okay, let me explain. For example, we have 5 price cloze entries and 1 price cloze exit, shifting the window bar by bar we are looking for a pattern in these 5 bars for the 6th. We have no stipulation in advance that the combination of these 5 bars looks the same. Now imagine that the combination looks the same every time, what is the neural network's answer? I don't think you need to answer that. Now further, in my case it turns out that combinations are always the same, but the length is different and it cannot be cut. Dependence of length is not the main for forward, but I think it matters too, so you cannot cut length. I had an idea to lengthen the ones that were shorter, but then they would lose their image, which is the original bet. Probably completely confused.....
If you don't cut back, then you need several neural networks, 100 by 200 and 250
Maxim, what do you feed to the input of a neural network? Wizard inputs increments, and you?
too them ) with different lag, I also want to add moments of distributions
and delay. Network with feedback. In fact, there are two networks.
But I've been lazy lately... probably because I've read too many books... I've read three 1000-page books in 1.5 weeks :D
I got out of it by adding a simple cycle to increase the length of the model, and now I get a good picture at any time. But the water forward still has the same 50/50 and to re-sort it, I'm now looking for methods...
portfolio is non-stationary, it doesn't go from sigma to sigma, it craps periodically... and then it recalculates and craps again
Say, if there is no global determinism like between some indices/shares, then trading a portfolio is like trading one symbol, but with additional costs
Or decompose one symbol, create a portfolio from it and trade one symbol))
If you don't cut it, then you need several neural networks, 100 by 200 and 250
Thanks for the advice, probably it would be more correct, and in the robot just put the signal from several networks instead of one...
Thanks for the advice, it would probably be more correct, but in the robot just put the signal from several networks instead of one ...
Why several? From the one with the length of the data submitted. You cannot get the answer from the second one with a different length.
So you can't get the ensemble. But you can get them one by one.
Why several? From the one with the length of the data you submitted. You can't get an answer from the second one with a different length.
That is, you can't get an ensemble. But one at a time, yes.
I misspoke, you're right. the logic will be the following, the model was built, the model length was determined and the trained network that matches the length of the current model was run. just as a result there will be several networks in the robot at once as well as models by length.
Well, the non-stationarity of the portfolio, it doesn't go from sigma to sigma, but cracks periodically... and then you recalculate it and it cracks again
Say, if there is no global determinism as between some indices/shares, then trading a portfolio is like trading 1 symbol, but with extra costs.
That's exactly right, but you have the advantage of giving the portfolio a form in which you can do your analysis every hour without having to wait for that form to appear on one pair. I will put it another way, for example I analyze just 10 bars from history to predict one bar in the future and a neural network will find hundreds of patterns from these 10 bars and I suggest to train a neural network on the forwards.