ESTIMATION OF TIMEVARYING AR(1)-PROCESSES WITH KERNEL METHODS

                      Jan Skowronski


In time series analysis stationarity is an important concept
which has received a lot of attention. However, time series
are often better represented as approximately locally
stationary. In this talk we consider AR(1)-processes where the
autoregressive parameters are timevarying.
 
Kernel methods based on local polynomial techniques have been
developed for estimators of the local covariance structure.
These local covariance estimators are then used in the
Yule-Walker equations when estimating the autoregressive
parameters.
 
A simulation study has been performed where we consider both
slow and abrupt changes in the autoregressive parameter
function. The Weighted Least Squares (WLS)-estimator of the
autoregressive parameter function is compared against the
methods above. Further we examine the mean square error and
the asymptotic mean square error.