Unified methods for detecting and handling outliers and structure changes in a univariate time series are proposed. The additive outlier and the involvement outlier are considered. Level shifts and variance changes are allowed, with level shift further classified as permanent level shift or transient level shift. A unified treatment is achieved by adopting a parametric approach to describe the exogenous disturbances. Under the parametric model, a disturbance's impact is measured by a parameter that can be estimated and tested under certain mild conditions. Two iterative procedures to detect and handle outlier and model changes when the number and time points of disturbances are unknown are recommended. These procedures, based on that of Chang (1982), consist of specification, estimation, detection, and removal cycles to handle one-by-one the most significant disturbances. The effectiveness of the methods is illustrated by analyzing 3 real data sets.
My Research