|
Abstract:
|
This presentation addresses modeling aspects of big time series data. Information criteria, such as BIC, are typically used for order identification in time series analysis. If no knowledge about the possible orders exists, the computation and minimization of an information criterion on a large enough grid of orders may help identify the optimal orders. However, computing the information criterion on a large grid of model orders becomes prohibitive for big time series. We propose to compute the information criterion only for a sample of orders and use kriging-based methods to emulate the information criterion on the rest of the grid. Then we use an EGO algorithm to identify the optimal orders. We focus mostly on ARMA time series models, but we also outline briefly how the method can be applied to more complex time series models, such as SARIMA. Both simulated and real big time series data are used to illustrate the method, showing that the model orders are accurately and efficiently identified.
|
