Design of the forecasting system related to structural changes is an important problem because structural changes induce abnormal forecast errors. In this thesis, we consider three types of structural changes in simple state-space model with a deterministic drift; level change, drift change and transient change.
We propose a Bayesian procedure to detect a change point and change type. For each point in a monitoring interval the posterior probabilities of each change type are derived and the minimal expected loss is obtained from a loss structure. Being based upon the minimum expected loss within the monitoring interval we choose the change type and change point. If any structural change is detected we reflect it to a forecast.
We examine the performance of the proposed procedure using simulation. The simplest loss is used for detection procedure. The results show small mis-identification rate and rapidity in identification of changes. For the better performance of this procedure it is needed to make a study of an appropriate loss-structure.
