Machine learning in trading: theory, models, practice and algo-trading - page 2657

New comment
 
Aleksey Nikolayev #:

Overlapping wavebreakers arean utterly disgusting sight.

Renate's right, if I understand him correctly.

The market is a non-stationary system.

Imagine that we have a TS on a moving average, except that we control the period of Mashka, for example, depending on volatility, etc... So the Mashka is no longer simple, but adaptive, the idea is: what if we do something similar with the rule?
 
mytarmailS #:
Renate's right, if I understand him correctly.

The market is a non-stationary system...

Imagine that we have a TS on a moving average, but we control the period of Mashka, for example, depending on volatility, etc... So the Mashka is no longer simple, but adaptive, the idea is: what if we do something similar with the rule?

Well, there is a rightness there only in the banal statement that there is no universal good solution for this problem.

To get rid of non-stationarity by constructing cleverly arranged signs is quite a normal idea, but not universal either, of course.

In my research I proceed from the standard assumption of piecewise stationarity, when the market has some stationary states between which it sometimes switches. Of course, it is not a universal approach either (non-stationarity may well be "floating").

 
Aleksey Nikolayev #:

Well, there is validity there except in the banal statement that there is no universally good solution to the problem.

To get rid of nonstationarity by constructing cleverly arranged signs is quite normal idea, but it is not universal, of course.

In my research I proceed from the standard assumption of piecewise stationarity, when the market has some stationary states between which it sometimes switches. Of course, it is not a universal approach either (non-stationarity may well be "floating").

The sadness of course is that it is not speed and acceleration)))))

Quite a normal approach to investigate. Stationary states are visible, modelled, transitions are also visible, but not modelled as stationary states. And floating non stationarity, it is still a complex substance, but noise always was and will be.

 
Aleksey Nikolayev #:

Well, there is validity there except in the banal assertion that there is no universally good solution to the problem at hand.

To get rid of nonstationarity by constructing cleverly arranged signs is quite a normal idea, but it is not universal, of course.

In my research I proceed from the standard assumption of piecewise stationarity, when the market has some stationary states between which it sometimes switches. Of course, it is not a universal approach either (non-stationarity may well be "floating").

It seems to me that the idea of piecewise stationarity is not working....

I argue: there is a place for delays, because stationarity is checked in a moving window, that is, we will detect it with a delay and we will know about the end of stationarity with a delay, it is like trading the crossing of MAs (always behind on the size of the window).


About switching, can we use HMM?
 
mytarmailS #:
I don't think the idea of piecewise stationarity is working.

I argue: there is a place for delay, because the stationarity is checked in a moving window, that is, we will detect it with a delay and we will know about the end of stationarity with a delay, it is like trading the crossing of MAs (always behind on the size of the window).


About switching, can we use HMM?

Reducing the lag leads to an increase in the false positive error rate, it's always a question of a trade-off between the two.

HMM is not quite suitable for me, because switching is not quite uniform, there are stuck states or too frequent switching. It is easier to assume that the switching moments are deterministic (and unknown).

 
Enumeration, enumeration... trigonometry, logarithms, moments of distributions, time-features... you can't guess what it should look like any other way

No luck with clustering yet.
 
Maxim Dmitrievsky #:
Enumeration, enumeration... trigonometry, logarithms, moments of distributions, time-features... you can't guess what it should look like any other way

Clustering's not working out yet.
What are you talking about?)
 
mytarmailS #:
What are you talking about?)

about looking for signs

 
Maxim Dmitrievsky #:

about looking for signs

there's nothing on the graph but increments and time. And derivatives don't give anything new. It is strange that clustering is stalled.

 
Valeriy Yastremskiy #:

there's nothing on the graph but increments and time. And derivatives don't give anything new. It's strange that clustering fails.

It links them to the target in a different way, in case of transformations. There could be different results. Clustering, in fact, and should not give anything :) you can certainly show off with the distribution of classes in clusters, but so far I do not get something sensible. That is, I take and correct the original zataset at the expense of clusters, supposedly the examples will be better distributed by classes. But it is in the current moment, and it is still difficult to predict the appearance of a cluster
1...265026512652265326542655265626572658265926602661266226632664...3591
New comment