Machine learning in trading: theory, models, practice and algo-trading - page 1978
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
It does work better, but it's a hell of a lot harder) if you add RL to it.
In general, conventional back propagation nets like mlp are not suitable for time series, not at all. At a minimum you need RNNFor stationary they are good) Simple logic for simple only works. For real ranks, a complex brain is needed.)
For stationary they are good) Simple logic for simple only works. For real series, you need a complex brain.)
Any nonstationary series can be represented as a sum of stationary ones of arbitrary length. But arbitrary length is a big problem in the forecasting problem.
Any non-stationary series can be represented as a sum of stationary ones of arbitrary length. Arbitrary length is a big problem in the prediction problem.
This is a big misconception.
For stationary they are good) Simple logic for simple only works. For real rows, you need a complex brain.)
depends on signal-to-noise ratio. At some point they stop working, because they do not take into account non-markness.
Roughly speaking, regularity disappears on noisy series (obvious cycles), but non-markness is preserved (if the process with memory). Normal mlp does not catch it, only RNN.
hence mlp, boosting, forest, etc. only for Markovian processes without memory.
Example with language: every language has a certain level of entropy, i.e. the alternation of words in the language. At a high level, speech becomes incoherent, for example, if there are a lot of parasitic words, or simply you are Peter Konov. Then you can catch only from the context, and for this you need the memory of past sentences (patterns).
For example, you read my sentence and don't know who Peter is and in what context I wrote it. You have no memory of past events and cannot relate them to the current wording, so you will draw the wrong conclusions.
depends on the signal-to-noise ratio. At some point, they stop working, because they do not take non-markness into account.
Roughly speaking, regularity disappears on noisy rows (explicit cycles), but non-markness is preserved (if the process with memory). Normal mlp does not catch it, only RNN.
hence mlp, boosting, forest, etc. only for Markovian processes without memory.
Example with language: every language has a certain level of entropy, i.e. the alternation of words in the language. At a high level, speech becomes incoherent, for example, if there are a lot of parasitic words, or simply you are Peter Konov. Then you can catch only from the context, and for this you need the memory of past sentences (patterns).
For example, you read my sentence and don't know who Peter is and in what context I wrote it. You have no memory of past events and cannot relate them to the current wording, so you will draw the wrong conclusions.
Signal/noise is of course determinative. In the case of strong noise, weak regularities will be lost, you just can't see them. But in the case of price series, noise is not created from the outside. Noise is fading regularities or weak, short ones, even if strong. But it doesn't change the essence. Regularities that can be detected and the rest is noise.
This is a big misconception.
Of course not for any in the full sense of the word any. White noise does not belong here, but we do not consider it either. We initially have a series made up of different regular sections, and they have different amplitudes and lengths, so we get a series in which there is noise and regularity.
Signal/noise is of course the determining factor. With a strong noise, weak regularities will be lost, they simply cannot be seen. But in the case of price series, noise is not created from the outside. Noise is fading regularities or weak, short ones, even if strong. But it doesn't change the essence. Regularities that can be detected and the rest is noise.
Well, if the noise is more than the signal, it is always an overfit or underfit (when using validation sampling). Because there are no stable patterns.
And when there is a lot of noise and few patterns, then try to isolate the signal.
it's really hard to understand why a sequence of noisy patterns contains a signal, but one pattern doesn't. We can simply increase the number of features (history being fed). But no, it doesn't work that way. Noise on noise produces noise. It requires more subtle context selection, it works there for some reason. Magic, in a word.
Well, if the noise is greater than the signal, it is always an overfit or underfit (when using validation sampling). Because there are no stable patterns.
And when there's a lot of noise and few patterns, you try to distinguish the signalWell, that's the point of all searches in all probabilistic topics, to isolate a pattern and catch the moment when it's gone. And to isolate is usually less problematic and costly).
:з
kind gentle polite))))
kindly affectionate polite))))
I'm doing a full logging, it'll tell you what he's doing.
this will give you an idea of what to improve