Machine learning in trading: theory, models, practice and algo-trading - page 2670

New comment
 

If you break the price into primitives - sinusoids, select the strongest sinusoids, you can see interesting patterns

Everything after the green vertical line is the future invisible to us, and all the curves after the line are forecasts.


Has anyone done anything in this direction?


I mean, you should enter at the beginning of a long, medium and fast trend.

 
mytarmailS #:

If you break down the price into primitives - sinusoids, select the strongest sinusoids, you can see interesting patterns

Everything after the green vertical line is the future invisible to us, and all the curves after the line are forecasts.


Has anyone done anything in that direction?


I mean you should enter at the beginning of a long, medium and fast trend.

sine wave extrapolation doesn't work.

 
Maxim Dmitrievsky #:

sinusoidal extrapolation does not work

Signals or factors here are mostly impulse, and DSP works with signals that are constant for some time, factors of influence. It also works with decaying ones, if you know the decay rate, and here it is with this as I understand the problem. It is possible to decompose in a certain time interval, it is possible to decompose a little later, but it is not yet learnt to correctly find the same sinusoids in the first and second time interval in systems where there is no constancy of signals. We can only assume that no new signals have appeared and none have faded. But this assumption is not correct for a market TS

 
Valeriy Yastremskiy #:

Signals or factors here are mostly impulse, and DSP works with signals that are constant for some time, factors of influence. It also works with decaying ones, if you know the decay rate, and I understand that this is the problem here. It is possible to decompose in a certain time interval, it is possible to decompose a little later, but it is not yet learnt to correctly find the same sinusoids in the first and second time interval in systems where there is no constancy of signals. We can only assume that no new signals have appeared and none have faded. But this assumption is not correct for a market TS

Yes, the search for a TS comes down to searching for similar repeating events, otherwise how. That's why we need filtering, not extrapolation
 
Maxim Dmitrievsky #:

sine wave extrapolation doesn't work.

I'm talking about a specific pattern with a filter, not throwing everything in a row and expecting a miracle
 
mytarmailS #:
I am talking about a specific pattern with a filter, not about throwing everything in a row and expecting a miracle
Do clustering on several components of the time series and see the predictions for each cluster. There will be random variation. This way you can go through the components for a long time, it's easier to zip the MO and put some filter on top of it

You can't do better than in my article. You can make different modifications, but this is the apogee of time series classification. You won't even find it in any articles. I invented it myself. It works great on synthetics, but for Forex you need to mess around with signs and labels and symbol search and other things

For example, I added markup on several symbols and retraining not from scratch, but iterative iteration. Speed increased, results not much. Now we need to work out the markup, get away from random markup.
 
In general, I think the power will soon allow more complete decompositions on close areas of a lot and analysis of changes in sinusoids or maybe other simple functions. So far, rough dabbling in this direction. In general, cardio is already in the clock, and even 5 years ago only on a computer was possible. Although it is not correct of course, cardio is still a permanent function of the system. But there is also enough noise there and the task of selecting the necessary signals automatically has not been fully solved yet.
 
Valeriy Yastremskiy #:
In general, I think the power will soon allow more complete decompositions on close areas of a lot and analysis of changes in sinusoids or maybe other simple functions. So far, rough dabbling in this direction. In general, cardio is already in the clock, and even 5 years ago only on a computer was possible. Although it is not correct of course, cardio is still a permanent function of the system. But there is also enough noise there and the task of selecting the necessary signals automatically has not been fully solved yet.
In parallel with this, the complexity and efficiency of markets will increase, so we will always be one step behind ) if earlier it was possible to write a primitive TS on the knee and it worked, now even the MO does not work.
 
mytarmailS #:
I'm talking about a specific pattern with a filter, not throwing everything in a row and expecting a miracle

PCA

the first component at the top, a few others at the bottom.

no extrapolation, forecasting, etc., real time mode...

The figure shows that a trend is when strong waves are in a certain order (pattern) those high-frequency, medium-term, long-term are in resonance....

Without decomposition, you just can't see it.


 
Maxim Dmitrievsky #:
In parallel with this, the complexity and efficiency of markets will increase, so we will always be one step behind ) if before it was possible to write a primitive TS on the knee and it worked, now even the MO does not survive

Nobody argues with this, it is possible to get closer to real time in the point of analysis, but not to overtake it of course. And the complexity of the system and analysis systems are parallel approximately.

1...266326642665266626672668266926702671267226732674267526762677...3498
New comment
Reason: