solving yule-walker equations
Hello
I want to estimate the parameters of an AutoRegressive model order p=5, and from netsearches I understand this entails
1) estimating the autocorrelationfunction up to and including lag 5
2) creating a correlationmatrix R from these, where r_ij = r_ji
3) solving the system R dot a = b
where a is unknown coefficients (the parameters), and b is r_1 - r_p.
I have substracted the mean from my series.
I compare my results from that of R (similar to Splus).
My estimations are perfect for the correlations, but my solution to the system is not. I use the method "gssitisol" from "A numerical library in Java for scientists and engineers", which supposedly should solve these kind of linear systems.
Is my System equation correct? Any other ideas on how to solve this?
Regards
JT
There is not to much to go on. I don't know what this "gssitisol" method is, and I also don't know if you wrote an implementation of it yourself from pseudo code or not...
But from my (very) limited knowledge of AR modelling (a topic in a class of 3 moths) I read that the Yule-Walker method can lead to poor estimates. The Burg抯 method is preferred (at least in this article):
http://bonsai.ims.u-tokyo.ac.jp/~mdehoon/publications/Annals.pdf
But I'm still a novice around here. I'm sure that if you're a bit more specific, and perhaps supply some code, there will surely be someone with far more math-skills here who can help you out.
Hope it helped.
Good luck.
Sorry, it is causal if 1-a1*z-...-ap*z^p <> 0 for all |z|<=1.
You can verify that the process is causal by testing this
with your estimated answer, but if a factor is found within
|z|<=1, I guess that no conclusion can be drawn since the
parameters might be wrong in the first place, or?
parza at 2007-7-15 16:07:19 >
