Data Mining On Time Series

With the development of modern information and technology, there are a lot of data about time series that contains many commercial information and need to be mined. Outlier mining is one of the most important of it. In conventional concept, outlier was often presumed to be noise or useless data and was removed in analysis. However, the noise data of one object may be the signal of another object, on this condition, the method may lead to the lost of some concealing signal. At present, outlier mining has attached a great importance in the field of time series analysis. Some important and influential work was done by Box and Tiao Abraham (1972), Barnett and Lewiss (1984), Chung Chen and Lon-Mu Liu (1993), Diaz(1991), Moculloch Tsay (1994), Cathy W.S (1997).But theirs work focus on two aspects which are (i) the purpose of outlier mining is to better estimate the parameters of model and the method to be adopted is remove these outliers but not consider the value in outliers itself, (ii) AR(p) model is mainly considered, or transform a ARMA(p,q) model to AR(p+q) model. Lon-Mu Liu etc (2001) illustrate an illustration using fast-food restaurant franchise data, which considered the commercial value in outlier data indeed. Based on the former work, in this paper, the motivation based on ARMA models with AO outlier mine the value of outlier. First, we introduce a general method on detecting outlier. Then we propose Gibbs sampling methods to mine outlier in the view of bayesian, for multiple sequence outliers, we propose a mended method and the simulation is done in computer. At last, we apply our method to real data and acquire better result.