An approach of solution to the change-point problem

Abstract

We developed a procedure to find change-points which seems both widely applicable and easy to implement on computer. The number of regimes is not needed a priori. Monte Carlo simulations showed that it is efficient also at low sample size. The approach is fully nonparametric and can be applied without knowing the statistical distribution of each subset. Confidence intervals are evaluated by a Monte Carlo simulation and this requires the latter distributions to be given explicitly. These are determined by Kolumogorov-Smirnov one sample goodness-of-fit test applied to all the different sets identified in the change-point analysis. The procedure is presented with a practical application to Gaussian and Poisson series.