Modular neural network applied to non-stationary time series

Modular artificial neural networks (MANN) have been used in the last years as clasification/forecasting machine, showing improved generalization capabilities that outperform those of single networks when the search space is stratified. Time Series data could be generated by many unknown and different sources and Modular Neural Networks, in particular Mixture of Experts models, are suitable for this time series where each expert is more capable to model some region in the input space and a gating network makes an intelligent selection of the expert that will model the specific pattern. Stochastical models for time series analysis are global models limited by the requirement of stationarity of the time series and normality and independence of the residuals. However, for most real world time series present behaviors such as heteroscedasticity, sudden burst of activity, or outliers. Such data are very common in finance, insurance, seismology and so on. In this paper we propose MANN models capable of dynamically adapt their architecture to non-stationary time series when the data is generated from several sources and is affected by the presence of outliers. Simulation results based on benchmark data sets are presented to support the proposed technique.