存在奇异值时AR模型的迭代稳健建模法

推广Abraham和Yatawara提出的检测时间序列中单个奇异值的评分检验法,用以检测AP(p)模型中多个相距大于2p+1的加性奇异值(A0).当检验到奇异值时,即用其预测值代替.用已经修改过的数据重新作阶数识别和参数估计,即可获得时间序列建模的稳健性.作为时序建模的典型应用,我们将该法用于谱估计.文中给出的计算机模拟例子表明,当时序中存在奇异值使通常的时序建模法的性能急剧恶化时,采用此方法,可以恢复到没有奇异值时该方法所能达到的性能水平.

AN ITERATIVE ROBUST MODEL-BUILDING PROCEDURE FOR AR PROCESSES IN THE PRESENCE OF ADDITIVE OUTLIERS

The procedure based on the score test for detection of single outlier in time series, discussed by Abraham and yatawara, is generalized to detect multi-outliers in AR(p) processes, which occur separately in time far more than 2p+l. Whenever an outlier is detected, it is replaced with its prediction, otherwise the data remain unchanged. The modified series is used in the ordinary model-building procedure, and robustizes the fitting procedure to the data. As an example, the procedure is used in robust autoregressive spectrum estimation, and the simulations verify the success of the procedure.