The RATS Software Forum

The attached zip has the examples and data sets from Commandeur & Koopman, An Introduction to State Space Time Series Analysis, Oxford University Press 2007. These are all examples of the use of the DLM instruction for analyzing state space models; in almost all cases, these are applications of structural time series models (local level or local trend with or without seasonality) which decompose a time series into the sum of two or more uncorrelated components. The examples are the most straightforward of any from the books that we have on the subject.

This book is available for purchase from Estima. For more information, see http://www.estima.com/textbook_commandeur.shtml.

Dear Tom:
In code of Chapter 8 - Norway & Finland data, forecasts, commandure & koopman(2007), there is a spelling error.
In finland code:
nonlin sigsqeps sigsqxi=0.0 sigsqzeta
dlm(a=at,c=ct,sv=sigsqeps,f=ft,sw=%diag(||sigsqxi,sigsqzeta||),exact,y=logfinland,$
method=bfgs,yhat=vhat,svhat=svhat) * 2008:1 xstates vstates
set forecast 1970:3 2008:1 = %if(t<=2003:1,%scalar(yhat),%scalar(xstates(t)))
set fvariance 1970:3 2008:1 = %if(t<=2003:1,%scalar(svhat),%scalar(vstates(t)))

vhat and yhat should be same.


Best Regard.
Hardmann

hardmann wrote:Dear Tom:
In code of Chapter 8 - Norway & Finland data, forecasts, commandure & koopman(2007), there is a spelling error.
In finland code:
nonlin sigsqeps sigsqxi=0.0 sigsqzeta
dlm(a=at,c=ct,sv=sigsqeps,f=ft,sw=%diag(||sigsqxi,sigsqzeta||),exact,y=logfinland,$
method=bfgs,yhat=vhat,svhat=svhat) * 2008:1 xstates vstates
set forecast 1970:3 2008:1 = %if(t<=2003:1,%scalar(yhat),%scalar(xstates(t)))
set fvariance 1970:3 2008:1 = %if(t<=2003:1,%scalar(svhat),%scalar(vstates(t)))

vhat and yhat should be same.


Best Regard.
Hardmann

Yes. It should read

dlm(a=at,c=ct,sv=sigsqeps,f=ft,sw=%diag(||sigsqxi,sigsqzeta||),exact,y=logfinland,$
method=bfgs,yhat=yhat,svhat=svhat) * 2008:1 xstates vstates

Dear Tom

I use STAMP 6.3 to estime the each model, the Log-Likelihood is very different from the one estimated from RATS, even is different from result from the Commandeur & Koopman book, so does AIC.
For example, as for the local determininstic level model for uk log KSI, the Log-Likelihood is 334.331 for STAMP, 63.3139 for RATS, 0.329757 for C&K book.
Why is these very different .

Best Regard
Hardmann

Have you checked whether the difference is due to the (log) integrating constants? RATS includes those, but the inclusion/exclusion has no effect on any of the estimates.

There are different ways to define AIC. @REGCRITS standardizes AIC by dividing by the number of observations---that doesn't change the ordering of models.