Machine learning in trading: theory, models, practice and algo-trading - page 1719

 
mytarmailS:

comment_15957283

What is the advantage of this approach? Why not use renges, or normalize with tick volume, or with average volatility at a certain time of day?

comment_15959144

Reminds me of speech recognition. Speech is recorded, converted to frequency, individual letters, words, etc. are recognized. The language often uses well-established phrases, you can guess the next word with a high degree of probability. What if this approach is transferred to the marketplace. Get a spectrum, try to distinguish patterns of "letters", "words", "phrases".


Maxim Dmitrievsky:

https://www.mql5.com/ru/forum/286022/page169#comment_15898101

https://www.mql5.com/ru/forum/286022/page169#comment_15898212

Any way to test on the Mersenne vortex?

Thinning from this article? If we exclude all but a specific clock from the series, wouldn't that be a transition to the daily TF?

 
Rorschach:

Is there a way to check on the Mersenne vortex?

Thinning from this article? Wouldn't it be a switch to a daily TF if we excluded all but a specific clock from the series?

I'll look at it later.

Yes, yes, daily cycles, but the increments can be taken with different lags

Roughly speaking, the logic is this: differentiate the series, thin out in any way (not necessarily fix intervals), look for linear dependencies. This is a universal thing, in my opinion, for finding dependencies.

 
Maxim Dmitrievsky:

I'll see later.

Yes, yes, daily cycles, but increments can be taken with different lags

Roughly speaking, the logic is as follows: differentiate the series, thin out in any way (not necessarily fix intervals), look for linear dependencies. This is a universal thing, in my opinion, for searching for dependencies.

From the early days, although I don't think I'd find anything now

 
Rorschach:

Any chance to check on the Mersenne vortex?

https://docs.scipy.org/doc/numpy-1.15.0/reference/generated/numpy.random.RandomState.html

generator = np.random.RandomState(0)

values = 0 + np.cumsum(generator.normal(0, 0.1, size=date_time.size))


numpy.random.RandomState — NumPy v1.15 Manual
  • docs.scipy.org
class seed=None¶ Container for the Mersenne Twister pseudo-random number generator. exposes a number of methods for generating random numbers drawn from a variety of probability distributions. In addition to the distribution-specific arguments, each method takes a keyword argument size that defaults to . If size is , then a single value is...
 
In fact, if we look at the private ACF on the increments of any SB, there are plenty of correlations. And if you look at the market increments, the correlations are weaker
 
Maxim Dmitrievsky:

That's funny. I couldn't predict the randomness. I still have a couple of thoughts, though.

I have a couple more ideas.

2. Calculate a prediction for initial gpsf and calculate prediction error (rsc gpsf-prediction). If worse than sko gpsch, then "tomorrow will be like today" is the best forecast ever.

 
Maxim Dmitrievsky:

I'll see later.

Yes, yes, daily cycles, but increments can be taken with different lags

Roughly speaking, the logic is as follows: differentiate the series, thin out in any way (not necessarily fix intervals), look for linear dependencies. This is a universal thing, in my opinion, for finding dependencies.

At first I thought it was about going to the daily timeframe, but since we differentiate first, it is not. And I'm a bit hung up, I don't see the "physical" sense of why it works. If viewed as a price flow, then we are limiting ourselves in observing price formation. If we view it as a signal, it is downsampling without prefiltering, which means aliasing and imaginary frequencies - signal distortion.

The most realistic explanation seems to be the influence of relic radiation) )) Then there must be diurnal and annual cycles.

 
Rorschach:

At first I thought it was about switching to a day timeframe, but since we differentiate first, it's not. And I'm a little hung up, I don't see the "physical" sense of why it works. If viewed as a price flow, then we are limiting ourselves in observing price formation. If viewed as a signal, then there is downsampling without pre-filtering, which is aliasing and imaginary frequencies - signal distortion.

The most realistic explanation looks like the influence of relic radiation) )) Then there must be diurnal and annual cycles.

thanks for the link )) i was just looking for something about natural cycles

the dependence of hourly increments with a lag of 24 from the previous one with the same lag, for example. I.e. looking at the previous one and predicting the current one

This works as long as the average increment deviates from zero (say, a sample of a year). There buying or selling accordingly. The articles have it all.

The reason for dividing by daily increments is clear. The American session is one process, the European one and so on. In other words, we regard the current European session as a continuation of yesterday's session, a specific hour or several hours, cutting out the rest.

Monthly cycles are also more or less clear. As for the rest, I do not know.

I tried to do the same for cointegrated (conditional) instruments, to improve cointegration, to take by hours. Better than the whole series, but uninspired.

 
It was a beautiful picture...
 
Maybe even trade sessions should be cut out separately and pasted together, without gaps, for clarity. For fun