Change point detection in a stochastic complexity framework

The authors present a method, inspired by stochastic complexity theory, for solving the change point detection problem for ARMA (autoregressive moving average) systems which are assumed to have a slow unstructured nondecaying drift after the change has occurred. The central idea is to apply the minimum description length method in the form of predictive stochastic complexity, which gives a way for selecting the best model among a given set of models. Therefore the change point detection problem is reduced to a model selection problem. Simulations that show that the approach exhibits good detection capabilities are included.<
>