Machine learning in trading: theory, models, practice and algo-trading - page 808
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
Totally agree, but then what can serve as a measure of stability?
I would trust if the results were adjusted only for the OOS, and then worked for the whole history of the instrument, but that seems almost unrealistic
and so far it's more of an art than a science.
Well, it's a multistep OOS (=crossvalidation).
In my opinion, no shamanism, pure science. We divide the story into 10 parts, develop it on one, test it on another. At the end we test it on the other 8, if it works in each of them, OOS is passed.I will definitely read the work you recommended.
Now I want to once again draw the attention of traders to one very important thing.
This thread is three years old. The result = 0. Why?
One of the reasons is the complete lack of understanding of what should be fed into the input of a neural network.
I stand by my conviction - to input BP increments. But of what series? Of those ones that are in numerous tick archives? OPEN OR CLOSE? On M1 or D1? No answer...
Considering that market processes are non-Markovian and cannot be predicted, the main struggle should be to transform BP - to try to reduce it to a Gaussian process.
Once again I ask the Man with a capital letter (I do not know who it will be) - take a tick archive from any brokerage house. Input the increments of this initial series to the input of the neural network. Show the results of the history test. Convert that BP to the Erlang flow of 1st order - the same, show the results, etc. Show the results here on the forum - people will help, if anything.
Only careful systematization of experiments can lead to success.
Well, it's a multi-step oos (=crossvalidation).
a more sophisticated fit, no less.
still have to check on the rest of the story afterwards
I will definitely read the work you recommended.
Now I want to once again draw the attention of traders to one very important thing.
This thread is three years old. The result = 0. Why?
One of the reasons is the complete lack of understanding of what should be fed into the input of a neural network.
I stand by my conviction - to input BP increments. But of what series? Of those ones that are in numerous tick archives? OPEN OR CLOSE? On M1 or D1? No answer...
Considering that market processes are non-Markovian and cannot be predicted, the main struggle should be to transform BP - to try to reduce it to a Gaussian process.
Once again I ask the Man with a capital letter (I do not know who it will be) - take a tick archive from any brokerage house. Input the increments of this initial series to the input of the neural network. Show the results of the history test. Convert that BP to the Erlang flow of 1st order - the same, show the results, etc. Show the results here on the forum - people will help, if anything.
Only careful systematization of experiments can lead to success.
I would feed just one parameter to neuronics input:
(high-open)/(open-low)
on periods from M30 and higherOne of the reasons is the complete lack of understanding of what should be fed to the input of the neural network.
I stand by my conviction - it is necessary to feed the BP increments to the input. But of what series? Of those ones that are in numerous tick archives? OPEN OR CLOSE? On M1 or D1? No answer...
If we consider that market processes are non-Markovian and cannot be predicted, then the main struggle should be to transform BP - trying to reduce it to a Gaussian process.
Hello Alexander!
Well, you are tired of your increments and reduction to the Gaussian process at the same time. I do not exclude that for your systems it is necessary, but generalization to all and everything is absolutely unfounded statements. In short, this is only your personal opinion, and no more than that. But do not impose it so insistently to all, as the truth in last instance.
I think that it is necessary to supply directly to the input of MO system a time series, normalized for a particular system. And I have already shown on the example of NS that it is both teachable and really works. Pictures of testing are given in this thread earlier.
However, this is probably not the only option, either.
I believe that the input to the MO system should be a direct time series, normalized for a particular system. And I have already shown on the example of NS that it is both trainable and really works. Pictures of testing are given in this thread earlier.
However, this is probably not the only option, either.
Yury, repeat again your pictures and experiments, because, by God, it's already not interesting to read the thread.
Yuri, repeat your pictures and experiments again, because, by golly, it is no longer interesting to read the branch.
That was before the New Year. There are many posts on the subject. Where do I find it now.))
In your branch recently showed a kind of good test-Graal (chart), a la Bollinger system (no NS). No increments. Only STO.
No implementation planned.
ZZY, though, there's something about NS in my blogs. There from the beginning of my classes, when I do not understand anything, to the moment when I decided on the use of NS.
Here.... new turn
And the CU is rowing.
Pound and on the upswing.
¶ what's he gonna bring ¶
# A new turn #
♪ And you can't tell ♪
Until you turn it, around the bend.....
I would feed just one parameter to the input of the neuron:
(high-open)/(open-low)
On periods of M30 and higherYou need to input what is the reason for price change, then any TS will work as it should. The price and everything that is built on it, including all kinds of distributions, are not the reason for the price, that is, by definition, cannot predict it. That's why you have these results..... IMHO...