We propose the penalized estimator with the smoothly clipped absolute deviation (SCAD) penalty for varying coefficient time series models, which in autoregressive models actually performs lag order selection. Theoretical properties are established. Some numerical examples are also presented.