A combined parametric and nonparametric approach to time series analysis. Motivated by coastal upwelling prediction (Thesis, TU München) (Q2726243)

The author started from the analysis of the coastal upwelling phenomenon using artificial neural networks as a suitable parametric model. In order to avoid the limitations of traditional neural techniques when analysing nonlinear and nonstationary time series, a new approach to time series analysis is proposed, called mixture of nonparametric segmented experts (MONSE). This approach is based on a combination of parametric and nonparametric techniques. It consists of the following four steps: nonparametric predictability test, heuristic segmentation of a given time series into non-overlapping parts with similar dynamics, parametric adaptation of the individual segments (using experts), and application. The MONSE approach is supposed to be applied to exploratory time series analysis and prediction, namely in the cases when hardly any knowledge about the time series under consideration is available.