Detecting Change Points in Time Series Using the Bayesian approach with Perfect Simulation
Degree GrantorUniversity of Canterbury
Degree NameMaster of Science
Many regression problems can be modelled as independent linear regressions on disjoint segments. The problem of interest is to find the number and location of the changepoints where the segments end, and to find the model order inside each segment. A new approach presented in Fearnhead (2005, 2006) is considered. This is a Bayesian approach using perfect simulation from the posterior distribution of the model. Some improvements to this algorithm are suggested: a method for selecting parameters for the prior distribution used, and an algorithm that eliminates a source of error found in Fearnhead (2005, 2006). This method is analysed by testing it on several simulations, with different model types and order. Three real datasets are then investigated; these are geological data, road safety data and medical data of preterm babies. Despite errors in certain situations, the algorithm is shown to be successful in many of the investigated cases, and an easy and efficient way of finding changepoints.