Time Series Model Selection via Adaptive Sparse Estimation

Model selection is a central agenda of autoregressive moving average (ARMA) modeling in time series data analysis. Recent advances in sparse estimation methods provide a fresh look at the time series model selection different from information criterion approaches. The adaptive LASSO method is paid attention in time series model selection due to its oracle property: the consistency of a set of non-zero parameters and its asymptotic normality. In spite of the solid theoretical property of adaptive LASSO method, this type is not a full-fledged method in time series analysis. This presentation will introduce a novel adaptive sparse method, the elastic net method, for time series model selection and investigate how the selection of initial estimates and tuning parameters in these adaptive sparse methods affects the performance of the time series model selection in various types of finite sample time series data. The performance will be assessed by examining the oracle property and the prediction accuracy. Comparison with other existing information criterion methods will be presented for both simulation studies and real data applications.