摘要
Most model selection methods for time series models with many predictors are devised for the stationary processes. We consider the problem of selecting higher-order autoregressive (AR) models whose integration orders can be positive or zero, and hence both stationary and non-stationary cases are included. Combining the strengths of AIC and BIC, we propose a two-stage information criterion (TSIC), and show that TSIC is asymptotically efficient in predicting integrated AR models when the underlying AR coefficients satisfy a wide range of conditions. We also conduct a simulation study to compare the performance of AIC, HQIC, BIC, TSIC, Lasso, the adaptive Lasso and the bridge criterion. Our study reveals that TSIC performs favorably compared to other methods in various scenarios.