|
PureBytes Links
Trading Reference Links
|
Hi,
Thanks Fred. With this formula we can compute directly DSP from AR coef.
I find very interresting things :
1- About how the autor estimate coef in NR. Estimator they mention for
autocorrelation calculation is the one not biased, and can lead to some
inaccurate result (and problem to result coefficient, so stability problem).
See 13.6.10 : http://www.library.cornell.edu/nr/bookfpdf/f13-6.pdf
Exemple:
x1=1.0; x2=1.1 et x3=1.0
r0=1.07, r1=1.1, r2=1.0
It should be impossible to have r(i)<r(i+1) and lead to some real problem
after (non positive matrix for some estimation method, inversion pbl,
singularity, pole outside unit circle maybe...).
Biased one is : Autocorr(i)=(1/N)sum(Yi*Yi+j,from j=1 to n-i)
1/(N-i) is replaced by 1/N.
Estimator is biased but variance is more weak than non biased one.
It has been show that the quadratic error (mean/variance) is less for the
biased one (Jenkin and Watts).
Author from NR use Burg algorytm to compute AR coef. I use Yuke Walker in
the posted AFL formula.
Do you try to compute this code in AFL : non-biased+burg ?
Is it accurate, fast ?
Other interessting section are :
3- subroutine fixrts -> can help fix the pole for stability so we can use
non-biased estimator and check after stability
4- (13.6.18) to remove the bias. For now i substract the mean and add it
back after prediction. I was thinking about polyfit to do that and take
spline trend in consideration.
If you got any code from "fixrts" and burg than you can share i will be
pleased to include it in AR model to add more option for the user, so he can
choose methode for compute coefficient and possible pole correction.
Maybe final stability will be improved so ARMA can be now compute too.
And last, do you make spectral comparaison between Burg and Yule Walker
Estimated DSP for quotes data ?
Cheers,
Mich.
----- Original Message -----
From: Fred
To: amibroker@xxxxxxxxxxxxxxx
Sent: Sunday, November 12, 2006 6:27 PM
Subject: [amibroker] Re: Polynomial Trendlines
If you look at the corresponding Fortran 77 link at NR you'll see the
graph that relates to output from calling the evlmem function.
_________________________________________________________________
Ten : profite de ton Messenger en illimité sur ton mobile !
http://mobile.live.fr/messenger/ten/
Content-Description: "AVG certification"
No virus found in this incoming message.
Checked by AVG Free Edition.
Version: 7.1.409 / Virus Database: 268.14.11/542 - Release Date: 11/20/2006
|