| |
 |
|
|
Science Forum Index » Statistics - Math Forum » a time series question: autocorrelation of very large lag
Page 1 of 1
|
| Author |
Message |
| MCI |
 Posted: Thu Dec 28, 2006 2:11 am |
|
|
|
Guest
|
Suppose I have 1000 daily observation of x, when i run regression of
x(t) - x(t-1) on x(t-1), r-squared is very close to 0; but when i run
x(t) - x(t-200) on x(t-200), r-squared can be as high as 0.65. (it
doesn't have to be exactly 200, just a large number around 200)
Anybody have an idea on how I can further explore this? (does this
imply that x is mean-reverting, but at a very slow speed?)
Thank a lot. Happy new year! |
|
|
| Back to top |
|
| vdebuen@gmail.com |
| Posted: Thu Dec 28, 2006 5:34 am |
|
|
|
Guest
|
MCI ha escrito:
Quote: Suppose I have 1000 daily observation of x, when i run regression of
x(t) - x(t-1) on x(t-1), r-squared is very close to 0; but when i run
x(t) - x(t-200) on x(t-200), r-squared can be as high as 0.65. (it
doesn't have to be exactly 200, just a large number around 200)
Anybody have an idea on how I can further explore this? (does this
imply that x is mean-reverting, but at a very slow speed?)
Thank a lot. Happy new year!
Could you have two or more big pulses than are separated by near 200
days? Time series that are sensible to social phenomena like
vacations, holidays and others structures, could show strange
correlations that analysts must capture with deterministic inputs.
Sometimes, a few days with anormal values may dominate the correlation
of 1000 normal values. |
|
|
| Back to top |
|
| Adam |
| Posted: Thu Dec 28, 2006 7:35 am |
|
|
|
Guest
|
"MCI" <330006@gmail.com> wrote in message
news:1167286264.424242.39630@f1g2000cwa.googlegroups.com...
Quote: Suppose I have 1000 daily observation of x, when i run regression of
x(t) - x(t-1) on x(t-1), r-squared is very close to 0; but when i run
x(t) - x(t-200) on x(t-200), r-squared can be as high as 0.65. (it
doesn't have to be exactly 200, just a large number around 200)
Anybody have an idea on how I can further explore this? (does this
imply that x is mean-reverting, but at a very slow speed?)
Thank a lot. Happy new year!
Have you tried plotting a graph of x against t? That would show you at a
glance if there is some unusual pattern to the data.
Adam |
|
|
| Back to top |
|
| Beliavsky |
| Posted: Thu Dec 28, 2006 10:26 am |
|
|
|
Guest
|
MCI wrote:
Quote: Suppose I have 1000 daily observation of x, when i run regression of
x(t) - x(t-1) on x(t-1), r-squared is very close to 0; but when i run
x(t) - x(t-200) on x(t-200), r-squared can be as high as 0.65. (it
doesn't have to be exactly 200, just a large number around 200)
The regression coefficient of x(t)-x(t-200) on x(t-200) is not the same
as the autocorrelation coefficient. Maybe you should compute and graph
the ACF of the original and differenced time series using a statistical
software program such as R. |
|
|
| Back to top |
|
| MCI |
| Posted: Thu Dec 28, 2006 10:27 am |
|
|
|
Guest
|
Thanks a lot for the help from you and Adam below.
When I scatter-plot x(t+200) - x(t) and x(t), it is very condensed near
a straight line,
so the high r-squared seems to be valid (not due to some outliers).
but the plots of x(t) Vs t and x(t+1) - x(t) Vs x(t) don't show any
visible pattern.
vdebuen@gmail.com wrote:
Quote: MCI ha escrito:
Suppose I have 1000 daily observation of x, when i run regression of
x(t) - x(t-1) on x(t-1), r-squared is very close to 0; but when i run
x(t) - x(t-200) on x(t-200), r-squared can be as high as 0.65. (it
doesn't have to be exactly 200, just a large number around 200)
Anybody have an idea on how I can further explore this? (does this
imply that x is mean-reverting, but at a very slow speed?)
Thank a lot. Happy new year!
Could you have two or more big pulses than are separated by near 200
days? Time series that are sensible to social phenomena like
vacations, holidays and others structures, could show strange
correlations that analysts must capture with deterministic inputs.
Sometimes, a few days with anormal values may dominate the correlation
of 1000 normal values. |
|
|
| Back to top |
|
| |
|
Page 1 of 1
All times are GMT - 5 Hours
The time now is Tue Oct 14, 2008 6:51 am
|
|