How to form the input values for the NS correctly. - page 17
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
This point, if you can expand on it, is the most confusing.
What rationing means in general. And how to normalize so that in the future the range will be in the 0-1 limit and we do not care about the values of the initial input non-normalized data.
On which sample should we normalize (all samples or just the current one)?
What kind (lined or s-type function).
If linear, the range in the future 0-1 may not be saved (a value greater than 1 will be found) and the network will not be trained on it.
If s-species, there is a saturation of the large ones and they will no longer be distinguishable for the network.
Is there a middle ground?
That's how we'll find her. We have to start with something first and then go through all the known ways to find the best one
Research the whole story, find the marginal range, use it for rationing and still with a margin
As far as I understand the aim of the branch is to improve network prediction by transforming input data, so we don't care how the network predicts...
Just the purpose of the thread is "lost":
- If the author has "favourite inputs" (e.g. returns - "classics of the genre"), "favourite" network architecture (yep :) ) and "teacher" (sentence from page 5), then the branch topic sounds: how to convert to range "activation functions" and, for me, please :), how to weed out "bad inputs". I.e. the given question may sound like this: if you stupidly, over the whole range, "scale" to [-1:1], is the learning/cross.../training error digestible and I am interested in a more "correct method"? I doubt it.
By the way, application of various "primitive" methods of "scaling" did not give a jump in error reduction in my!!!
- If the author is in search of "good inputs", then, firstly, ALL at once refer to the "intimate" and the topic stalls, and secondly - everyone has his own vision, including what and how to teach - there is already a muddle. Only to experiment and, for me, please :), how to screen out the "bad inputs"
`
Otherwise - "disarray and vacillation" and another branch is stalled. Therefore, we need specialized forums with specialized branches with the most severe moderating and monitoring of the topic of "performances", which ... are also "stalled out", but for a different reason. (Not in front of klot will I say).
`
ZZY. For example, TradingSolution has such a "utility" - "Correlation Analysis" with which, ostensibly, one can evaluate whether the "Teacher" (in TS terms - "Optimal Signal") "correlates" with the input of interest. But I still will not believe that such "dumb correlation" between OHLS and "Teacher" can give values greater than 0.5 even to "really meaningful inputs". Unless the "all you can" is smoothed out. And when the teacher is hm ... discrete, as "inputs from ZigZag", I, for example, had "errors" that were out of bounds. Personally I was interested only in it - "Data preprocessing" (chapter 7 in Ezhov, but there at first one formulas, and secondly -... I don't really believe in applicability of all mentioned there for such "noisy"/discrete data, and I can't manage to write "subroutines" :( ) and it seemed to me, that this branch is devoted precisely to it, but...
SZY. I think of all well - probably you now sit and pore over concrete implementations, and in a branch you splash out only "theoretical difficulties". :)
Guys, you've gone the wrong way......)))))
Research the whole story, find the limit range, use it for rationing and still with a margin
Yes.you already said that and even gave a picture explaining it. It's all clear. It's just that then this question came up with the future and so in short I'm a bit lost and pondering. Have you ever tried to use s-functions for the inputs themselves? Or only linearly?
Just the purpose of the branch is 'lost':
Let the purpose of the branch be as everyone sees it for themselves. I will not be offended. :) The more ideas, the better the vision of the problem.
But perhaps you have very wisely pointed out that when it comes down to it :
Then, firstly, ALL immediately refer to "intimacy" and the topic goes dead, and secondly - everyone has a different vision,
So there's only one thing left to do.
Only to experiment.
And in the branch let people express their thoughts and experiences.I'm sure that the next it will be very interesting.
Here we go. Wrote and wrote, then decided to "fix it" and it got erased. :(
Anyway:
1. In the window (whose size theoretically!!! also affects "profitability"), which moves when a new bar appears:
For me, none of the three methods, either on the whole sample or "by the window" brought any fundamental improvement.
From this I concluded that the main thing is inputs ... adequate output. (But this is, unfortunately, another tale. ;( )
But "Joint rationing: whitening the inputs" (p.132), may be more curious.
2. In reply to "Yes.you already said that and even gave a picture explaining it.It's all clear. It's just that then this question with the future came up... " I asked " Was itgood in the past?"
'
And wanted to add that on the spider, next too was doing 'research' on different rationing in TS.
'
ZY=IMHO. And rationing is better in ... Excel, by setting NS4 on a sample, for example, and building, on results, a graph "Error=f(type of rationing, size of a window)". Then the questions will disappear.
My IMHO, the zigzag as inputs to the NS is a useless case, and as a compression of information, too. It shows peaks, but in no way reflects the dynamics in between. Especially since it reacts to almost any spike, so, again, IMHO, screw it.
history doesn't react to a spike anymore :-)
I was talking about history - short term of course - but history
and i was saying that there are indicators that are good to find these spikes in the reversals
the question is what to teach? the nearest history !
1 pivot points - i.e. change in trend
2 or something else ?