不同类型识别变量的自回归模型异常值探测的Bayes方法

Bayesian Methods for Outliers Detection in Autoregressive Model Based on Different Types of Classification Variables

讨论基于自回归模型(AR模型)的时间序列数据中异常值探测的Bayes方法。该方法针对自回归模型引入不同类型的识别变量,通过比较这些识别变量的后验概率值与事先给定的阈值来进行异常值定位;基于Gibbs抽样算法,提出识别变量后验概率值的计算方法和异常值的估算方法;进行了大量的模拟试验并把该方法应用于卫星钟差实测数据的异常值探测,结果表明,该方法对于解决时间序列数据中在同一时刻或不同时刻出现加性异常值或革新异常值的探测问题是可行的和有效的。

ABayesian procedure for outlier detection in time series is discussed. The main idea of this method is introducing different types of classification variables into autoregressive model. Then outliers can be detected by comparing the posterior probabilities of these classification variables with a given threshold. Besides, a procedure for computing the posterior probabilities of classification variables and obtaining the estimates of outliers is designed based on Gibbs sampling. A large number of simulation experiments and an experiment of real clock error data are carried out. It is shown that the new procedure is applicable to detect additive and innovational outliers occurring at the same time or not in time series.