How do you practically assess the contribution of a "specific" input to the NS? - page 4
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
We can also propose the following variant: for simplicity we take NS with three inputs.
And apply all 20 inputs and let the optimizer find the best combination of 3 inputs on its own by criterion, for example, a forward run for minimal drawdown.
Something like this.
We can also propose the following variant: for simplicity we take NS with three inputs.
And apply all 20 inputs and let the optimizer find the best combination of 3 inputs on its own by criterion, for example, a forward run for minimal drawdown.
Something like this.
Not quite Friday, but ...
There is an NS, any NS, there is an input A={A1, A2, .... A20}. Train the NS and get a satisfying result. How do we practically evaluate the contribution of each element of input A1, A2, ... A20 to this result?
The options off the top of my head are:
1) Somehow sum up and calculate all weights with which the element passes through the network. I'm not quite clear how to do it, I would have to immerse myself in the network operation and calculate somehow some coefficients, etc.
2) Try to "zero out" somehow, or e.g. reverse an element of input vector and see how it affects the final result. So far I've settled on it.
But before realizing this second variant I decided to ask my advice. Who may have been thinking on this subject longer than me? Maybe someone can advise me on some useful book-article?
I propose to write an indicator and run it in a separate window.
Indicator lines would make for very interesting cognitive observations.
Indicator lines may be: neuron adder outputs; neuron outputs after non-linear transducers; possibly committee outputs, etc. Everything depends only on your desires and fantasies.
Such visibility will help to "penetrate" into this black box and understand how it all happens/works there.
That's for sure. Everyone has a choice: read books or develop an unrestrained flight of fancy.
Just for the record, this is the method described in the books.
Perhaps not in the form I have submitted here, but essentially correct.
The degree of influence of each input is almost impossible to realistically estimate .
I don't know. Relative to the other inputs, it's fine. Only the inputs need to be normalised.
So we take the predicted outputs as a benchmark and for each input for all patterns we calculate the RMS error for some very small shift of a certain input.
Applying a solid evidence-based approach outside its econometric context raises childish questions.
Doing a regression:
Profit = s(1) * A0 + ... s(n) * A(n)
We estimate the coefficients of this regression.
Immediately we get
probability that a specific coefficient is equal to zero - delete such an input
probability of all coefficients being equal to zero taken together
using ellipses we get correlation coefficients
perform the test for redundant inputs
perform a test on missing inputs
test for the stability of coefficient values (evaluate the randomness of their change)
An intelligent man came and gave an adult answer to my childish question) Thanks to that. Not only regression and NS are not quite the same thing, but the proposed option is at least not simpler. We estimate, get, conduct, conduct, conduct... And it is not clear how to interpret results obtained in quite a different system. Is MACD good or bad? Can one TS use it, while the other does not?
We can also suggest the following variant: we take an NS with three inputs for simplicity.
We feed all 20 inputs and let the optimizer find the best combination of 3 inputs on its own by criterion, for example, a forward run for minimal drawdown.
Something like this.
I did exactly the same thing, but did not take any inputs and created combinations of them. I excluded inputs and some combinations of them and watched the result - which is the same thing. Switching on, switching off - what's the difference? Due to the specifics of the implementation, I found it more convenient to exclude.
That's right. Everyone has a choice: read a book or develop an unrestrained flight of fancy.
Then again, I've even asked about the article books as well. No one suggested anything on the subject and neither did you. Go to the library of science and technology and pray for econometrics, the only non-science?) Although I really don't mind books, if I browse them in the toilet with the aim of education or culture, they are of very little practical use and have no ready-made solutions, because they are written by fundamentalist theorists or by unsuccessful applied scientists. And no matter how much you read them, without "unbridled imagination" they are of no practical use.
I don't know. Relative to the other inputs, it's fine. Only the inputs should be normalized.
So we take the predicted outputs as a benchmark and for each input for all patterns, we calculate the RMS error for some very small shift of a certain input.
By the way, NS is also a regression. The same dependence of the current countdown on previous countdowns. But that's not the point.
What faa suggests is applicable to linear regression, and neural network is a non-linear regression.
By the way, NS is also a regression.