I looked into why the coefficients differ in Gabor's example and it's because
the example mistakenly switched the order of x and y. The syntax is
roll::roll_lm(x, y), which is the same as fastLm, but the example accidentally
switched them. After correcting the example, the coefficients are all e
> jeremiah rounds
> on Thu, 21 Jul 2016 13:56:17 -0700 writes:
> I appreciate the timing, so much so I changed the code to show the issue.
> It is a problem of scale.
> roll_lm probably has a heavy start-up cost but otherwise completely
> out-performs those other vers
Hi Jeremiah: I think I wasn't that clear. I'm not suggesting the kalman
filter to deal with time varying coefficients. As Roy pointed out, one can
use the kalman filter to do regular regression where one "sees" a new data
point as each time unit passes. It can be assumed that the coefficients do
n
This may be useful:
Sven Hammarling and Craig Lucas
"Updating the QR factorization and the least squares problem"
http://eprints.ma.man.ac.uk/1192/01/covered/MIMS_ep2008_111.pdf
http://www.maths.manchester.ac.uk/~clucas/updating/
2016-07-21 20:02 GMT+02:00 jeremiah rounds :
> Hi,
>
> A not unusua
I agree that when appropriate Kalman Filter/Smoothing the higher-quality
way to go about estimating a time-varying coefficient (given that is what
they do), and I have noted that both the R package "dlm" and the function
"StructTS" handle these problems quickly. I am working on that in
parallel.
I have no idea which method produces the fastest results, but the package KFAS
has a function to do recursive regressions using the Kalman filter. One
difference is that it is not, as far as a I can telll, a moving window (so past
data are being dropped), just a recursively computed regressio
Hi Jermiah: another possibly faster way would be to use a kalman filtering
framework. I forget the details but duncan and horne have a paper which
shows how a regression can be re-computed each time a new data point is
added .I
forget if they handle taking one off of the back also which is what you
I appreciate the timing, so much so I changed the code to show the issue.
It is a problem of scale.
roll_lm probably has a heavy start-up cost but otherwise completely
out-performs those other versions at scale. I suspect you are timing the
nearly constant time start-up cost in small data. I
I would be careful about making assumptions regarding what is faster.
Performance tends to be nonintuitive.
When I ran rollapply/lm, rollapply/fastLm and roll_lm on the example
you provided rollapply/fastLm was three times faster than roll_lm. Of
course this could change with data of different di
Thanks all. roll::roll_lm was essentially what I wanted. I think maybe
I would prefer it to have options to return a few more things, but it is
the coefficients, and the remaining statistics you might want can be
calculated fast enough from there.
On Thu, Jul 21, 2016 at 12:36 PM, Achim Zeile
Jeremiah,
for this purpose there are the "roll" and "RcppRoll" packages. Both use
Rcpp and the former also provides rolling lm models. The latter has a
generic interface that let's you define your own function.
One thing to pay attention to, though, is the numerical reliability.
Especially o
Just replacing lm with a faster version would speed it up. Try lm.fit
or even faster is fastLm in the RcppArmadillo package.
On Thu, Jul 21, 2016 at 2:02 PM, jeremiah rounds
wrote:
> Hi,
>
> A not unusual task is performing a multiple regression in a rolling window
> on a time-series.A stand
Hi,
A not unusual task is performing a multiple regression in a rolling window
on a time-series.A standard piece of advice for doing in R is something
like the code that follows at the end of the email. I am currently using
an "embed" variant of that code and that piece of advice is out there
13 matches
Mail list logo
