Multivariate time series model discovery with similarity-based neuro-fuzzy networks and genetic algorithms

This paper studies the properties of a hybrid technique for model discovery in multivariate time series, using similarity based hybrid neuro-fuzzy neural networks and genetic algorithms. This method discovers dependency patterns relating future values of a target series with past values of all examined series, and also constructs a prediction function. It accepts a mixture of numeric and non-numeric variables, fuzzy information, and missing values. Experiments were made with a real multivariate time series for studying the model discovery ability and the influence of missing values. Results show that the method is very robust, discovers relevant interdependencies, gives accurate predictions and is tolerant to considerable proportions of missing information.