Machine learning in trading: theory, models, practice and algo-trading - page 1594

 

Everyone's so smart, it's creepy.


But maybe someone will condescend from the height of his ego and explain to the poor dekhanin the commonplace truths?

let's assume that we have found that some series over 100 bars is conditionally stationary, i.e. we have a certain window of 100 bars

on the next bar we shift the window and see that the series is stationary again for 100 bars and then again and again... but on the 17th bar after the "stationarity detection" we see that the series has ceased to be stationary, i.e. "suddenly" it has lost its stationarity

However, we can suppose that if we had taken the series with the window length not of 100 bars but of 118 bars (100+17+1), this series would still be stationary

the van's question: why is it that within these 17 bars we cannot trade from MO+SCO to MO, or even more so from MO+-2*SCO to MO?

the question to usbe tu: "by whom" should stationarity be evaluated - by the series of 100 bars (on the moving window) or by the series of 117 bars (17 bars from "stationarity detection" to "stationarity end" + 100 bars backwards)?

 
Boris:

Everyone's so smart, it's creepy.


But maybe someone will condescend from the height of his ego and explain to the poor dekhanin the commonplace truths?

let's assume that we have found that some series over 100 bars is conditionally stationary, i.e. we have a certain window of 100 bars

on the next bar we shift the window and see that the series is stationary again for 100 bars and then again and again... but on the 17th bar after the "stationarity detection" we see that the series has ceased to be stationary, i.e. "suddenly" it has lost its stationarity

However, we can suppose that if we had taken the series with the window length not of 100 bars but of 118 bars (100+17+1), this series would still be stationary

the van question: why is it that within these 17 bars we cannot trade from MO+SCO to MO, or even more so from MO+-2*SCO to MO?

The question is "by whom" should stationarity be estimated - by the series of 100 bars (on the moving window) or by the series of 117 bars (17 bars from "stationarity detection" to "the end of stationarity" + 100 bars back)?

1. You can. But more often than not, trading on a stationary series is about opening a trade on the upper or lower boundary of the Canadian to return to MO or the opposite boundary. If stationarity is violated - your position on the boundary will go negative due to expanding variance. The loss will outweigh all accumulated profit.

2. Googling the question of sufficiency of the volume of sampling. As far as I remember, it all depends on the distribution function

 
Dmitry:

1. You can. But more often than not, trading on a stationary series is about opening a trade on the upper or lower boundary of the Canadian to return to the MO or the opposite boundary. If stationarity is broken - your position on the boundary will go negative due to expanding variance. The loss will outweigh all accumulated profit.

2. Google the issue of sufficiency of the volume of sampling. As far as I remember, it all depends on the distribution function

1. I know I can, but someone here recently argued that stationary series are not predictable

As for the loss, strictly speaking, it's not, but if you find a "way to combat" it, you can dramatically increase the profitability

2. Google will not help here from the word "at all". The question is very simple. Which series to use to evaluate the stationarity - a constant length or lengthening series?

and each of them can have its own, only inherent "stationarity"

 
Dmitry:

Where is the source of "big data"?

Is there a database?

You have to build it yourself, how else can you do it? Put on a vds\vps writer and enjoy life, after a year or so.

You can, of course, and buy, but it will be expensive and definitely not all that will pop into your head. And the whole point is in the "not yet trampled" data, not something that everyone has and everyone uses for the same purpose.

 
Boris:

1. I know I can, but someone here recently argued that stationary series are not predictable

as for the loss, strictly speaking, they are not, but if you find a "way to combat" it, you can dramatically increase the profitability

2. Google will not help here from the word "at all". The question is very simple. Which series to use to evaluate the stationarity - a constant length or lengthening series?

And each of them may have their own inherent "stationarity".

1. Well, you should write to the person who argued that.

1а. Have you found a "way to struggle"? If so, share it with Alexander A_K.

2. Once again, sufficiency or optimal sample size. Google

 
Dimitri:

1. Well, you should write to the person who claimed it.

1а. Did you find a "way to fight"? If so, share it with Alexander A_K.

2. Once again, sufficiency or optimal sample size. Google

1a. who is this?
 

Here is a graph

the process you recognize "out of 1000" ))) the line with a step in the center is MO, a step when the process is no longer stationary on the shifting window, although we see signs of this even earlier

the blue 5th row is the changing MO (it depends on how you count the length of the row)

2 rows above and below are plus or minus 1-2 RMS from the MO

And what do we see? The process returns anyway and does not care that it is "non-stationary" anymore


Here is another graph



Of course, it can also be like this

All this time we are "stationary"

then "non-stationarity", a step and voila

we can wait for the return again

 
Boris:

Here is a graph

the process you recognize "from 1000" ))) the line with a step in the center is the MO, the step when the process is no longer stationary at the shifting window, though we see signs of it even earlier

the blue 5th row is the changing MO (it depends on how you count the length of the row)

2 rows above and below are plus or minus 1-2 RMS from the MO

And what do we see? The process returns anyway and does not care that it is "non-stationary" anymore


Here is another graph



Of course, it can also be like this

All this time we are "stationary".

then "unsteady", a step and voila.

we can wait for the return again

There are also returns for non-stationary processes (e.g., SB)

Stationarity (by definition) is:

1) constancy of expectation

2) dispersion constancy

3) ACF dependence on the time difference only

How exactly do you check all this?

 
Aleksey Nikolayev:

Returns also occur for non-stationary processes (e.g., SB)

Stationarity (by definition) is:

1) constancy of expectation

2) dispersion constancy

3) ACF dependence on the time difference only

How exactly is all this checked?

adf_test in R
 
Boris:

here is a graph

you will recognize the process "from 1000" ))) the line with a step in the center is the MO, a step when the process is no longer stationary on the shifting window, although we see signs of it even earlier

the blue 5th row is the changing MO (it depends on how you count the length of the row)

2 rows at the top and bottom are plus or minus 1-2 RMS from the MO

And what do we see? The process returns anyway and does not care about the fact that it is "non-stationary" anymore


here is another graph



Of course, it can be like this

all this time we are "stationary"

then "unsteady", a step, and voila.

we can wait for the return again.

You don't need to trade changing modes, you need to change strategies when they change. If it's scalping, there will be hundreds of trades for each one. The task is to switch strategy in time, i.e. to determine the mode change as early as possible, or even predict.

If you solve this problem, the Grail is definitely in your pocket