Search

We indicate how to handle a large database consisting of nonlinear time series, applying different nonlinear modelling techniques to this kind of times series. Nowadays in the current references there is no explicit guide of how to manipulate data from nonlinear time series; however, there are approaches that can be taken account. To this end we studied a monthly database corresponding to South Oscillation Index (SOI) and between the years 1866 to 2006. It explains how there must manipulated this database whose data possess nonlinear characteristic, which will be used to do forecasts several steps ahead. Two standard tests to this database were applied, the Average Mutual Information (AMI) and the False Nearest Neighbours (FNN). The optimal spacing of the information was obtained as well as the number of values backward necessary to predict values towards the future. Then, several models were designed of artificial neural nets (ANN), with different learning rules, function of transfer, elements of process (or neurons) in the hidden layer, etc., that allowed to do forecasting of up to 20 steps ahead. The best networks were those that possessed the rules of learning called extDBD and Delta-Rule, and sigmoid as well as hyperbolic tangent as function of transfer. The type of used network was one of feedforward multilayer perceptron and trained by means of backpropagation technique. Networks were proved by one, two hidden layers and without any hidden layer. The best model that was obtained it turned out to be one that consisted with an alone hidden layer. [ABSTRACT FROM AUTHOR]

Time series prediction is an important research problem due to its implications in engineering, economics, finance and social sciences. An important topic about this problematic is the development of new models and its comparison with previous approaches in terms of forecast accuracy. Recently, support vector machines (SVM) have been used for time series prediction, but the reported experiences are limited and there are some problems related to its specification. The aim of this paper is to propose a novel technique for estimating some constants of the SVM usually fixed empirically by the modeler. The proposed technique is used to estimate several SVM with the aim of forecast five benchmark time series; the obtained results are compared with the statistics reported in other papers. The proposed method allow us to obtain competitive SVM for the time series forecasted in comparison with the results obtained using other most traditional models. [ABSTRACT FROM AUTHOR]