Inference for nonstationary time series by adaptive weights smoothing

V. Spokoiny
Weierstrass Institute and Humboldt University, Berlin, Germany

A  nonstationary process can be modelled by an autoregressive
equation with time dependent coefficients. The Adaptive Weights
Smoothing (AWS) procedure introduced in Polzehl and Spokoiny
(2000) can be effectively applied to analysis of such models. The
method is fully adaptive and it allows to proceed in a unified way
for change-point and smooth transition models. It also leads to the
classical estimator of the autoregressive parameters under time
homogeneity. The performance of the method is illustrated by the
simulated examples and applications to exchange rate and EEG data
sets.

Reference

J. Polzehl and V. Spokoiny (2000).
Adaptive weights smoothing with applications to image segmentation.
J. of Royal Stat. Soc., 62, Series B, 335-354.