Time-Series Prediction Using Self-Organising Mixture Autoregressive Network

In the past few years, various variants of the self-organising map (SOM) have been proposed to extend its ability for modelling timeseries or temporal sequence. Most of them, however, have little connection to, or are over-simplified, autoregressive (AR) models. In this paper, a new extension termed, self-organising mixture autoregressive (SOMAR) network is proposed to topologically cluster time-series segments into underlying generating AR models. It uses autocorrelation values as the similarity measure between the model and the time-series segments. Such networks can be used for modelling nonstationary timeseries. Experiments on predicting artificial time-series (Mackey-Glass) and real-world data (foreign exchange rates) are presented and results show that the proposed SOMAR network is a viable and superior to other SOM-based approaches.