Machine learning in trading: theory, models, practice and algo-trading - page 358

New comment
 
Yuriy Asaulenko:
I would like to be more specific. I can give two opposite answers.)


Why two answers?

General rule of thumb:

  • On stationary series the prediction is the same as on historical data
  • Non-stationary series are not predicted without prior effort
There is ARMA, there is ARIMA, which can predict non-stationary series, but this kind of non-stationarity is very rare. There's ARCH, there's a bunch of different GARCHs, all of which take into account different kinds of nonstationarity for the sake of being able to predict future motion.


Can NS predict non-stationary series? If yes, what types of nonstationary series?

 
Yuriy Asaulenko:

A new maximum (probably the minimum) is followed by a new maximum - yes, I also went through this, the graphs are all familiar. I simulate - and there's nothing there - nothing. Maybe you will be lucky.


This is in the case of unstable (antipersistent series).

And in the case of persistent (steady), a new maximum is followed by a new maximum.

The problem is that the MAKA is heavily redrawn at a low period, ie it can not apply. If we take it for n-bars back the signal will be already missed.

 
Maxim Dmitrievsky:


The problem is that the MA is heavily overdrawn at a low period, i.e. it cannot be applied. If we take it for n-bars backward the signal will be already missed.

the other day I was playing with MAs (not simple ones, but gold ones)) - 3rd order filters. The 12-MAs have a group delay of 4 minutes. Let's not even talk about the EMA and other standard ones - the lag is off the scale.

In general, it is necessary to get away from the MA to the regression line. But the calculations are very delayed there. If on 1 minute with allowance for ticks, it will be fatal.

 
SanSanych Fomenko:
Does anyone know the answer to the question: how do NSs treat nonstationary inputs?
The neural network doesn't care if the input is stationary, non-stationary, or not timed at all. It makes no difference. Especially when it comes to classification
 
Vladimir Perervenko:
The neural network doesn't care - stationary, non-stationary, or no time series at all. It makes no difference. Especially when it comes to classification.
That's what I meant as one of the answers.)
 
Vladimir Perervenko:
The neural network doesn't care if it's stationary, non-stationary, or not a timeseries at all. It makes no difference. Especially when it comes to classification
It is very desirable for inputs and outputs to be limited by value area.
 
Combinator:
it is very desirable that inputs and outputs be limited by the domain of values.
Such issues should be solved even before entering the NS. NS, as a rule, does not eat raw data.
 
Vladimir Perervenko:
The neural network doesn't care - stationary, non-stationary, or no time series at all. It makes no difference. Especially when it comes to classification.

Then it is a question of retraining in all its glory
 

I don't even know if I need to do any better, 20,000% in 2.5 months at opening prices on 5 minutes if I'm lucky... you throw in $1k and pre-order a Bentley. If you're unlucky, it's no big loss )


 
Maxim Dmitrievsky:

I don't even know if I need to do any better, 20,000% in 2.5 months at opening prices on 5 minutes if I'm lucky... you throw in $1k and pre-order a Bentley. If you're not lucky, no big loss.)

Spell it out.)) I want, if not a Bentley, then at least a Peugeot with an automatic.)
1...351352353354355356357358359360361362363364365...3494
New comment
Reason: