Modeling and forecasting streaming data has fundamental importance in many real world applications. In this paper, we present an online model selection technique that can be used to model non-stationary time series in a sequential manner. Multi-state autoregressive (AR) model is used to describe non-stationary time series, and a dynamic algorithm is applied to learn the states at each time step. The proposed technique estimates a candidate AR filter from the most recent data points at every time step, and checks whether starting a new state significantly decreases prediction error or not. To that end, a time-varying threshold is compared with the reduction in the prediction error caused by postulating a new AR filter. The threshold is calculated by sampling and clustering uniformly distributed stable AR filters. Numerical simulations show that the proposed algorithm accurately estimates the state transitions with a small delay.