Random Walk Smooth Transition Autoregressive Models

This paper extends the family of smooth transition autoregressive (STAR) models by proposing a specification in which the autoregressive parameters follow random walks. The random walks in the parameters can capture structural change within a regime switching framework, but in contrast to the time varying STAR (TV-STAR) speciifcation recently introduced by Lundbergh et al (2003), structural change in our random walk STAR (RW-STAR) setting follows a stochastic process rather than a deterministic function of time. We suggest tests for RW-STAR behaviour and study the performance of RW-STARmodels in an empirical setting. The out-of sample forecasting performance of our RW-STAR models is encouraging - better than AR, LSTAR and TV-STAR specifications with respect to point forecasts and on a par with TV-STAR speciÞcations with respect to forecast density evaluations.