Optimization of Neuro-Coefficient Smooth Transition Autoregressive Models Using Differential Evolution

This paper presents a procedure for parameter estimation of the neuro-coefficient smooth transition autoregressive model, substituting the combination of grid search and local search of the original proposal of Medeiros and Veiga (2005, IEEE Trans. NN, 16(1):97-113) with a differential evolution scheme. The purpose of this novel fitting procedure is to obtain more accurate models under preservation of the most important model characteristics. These are, firstly, that the models are built using an iterative approach based on statistical tests, and therewith have a mathematically sound construction procedure. And secondly, that the models are interpretable in terms of fuzzy rules. The proposed procedure has been tested empirically by applying it to different real-world time series. The results indicate that, in terms of accuracy, significantly improved models can be achieved, so that accuracy of the resulting models is comparable to other standard time series forecasting methods.