In this work, an algorithm for identifying time series models is proposed. The strategy is based on Partial Mutual Information Criterion (PMI), which considers not only linear but also non-linear relations between variables under study. For calculating the PMI criterion, it is necessary to approximate marginal and joint probability densities, as well as conditional expected values. In this work, these operators are estimated using the city-block distance function and product multivariate kernel estimators. The algorithm is applied for identifying time series linear models and for selecting inputs for a non-linear model based on neural networks. The neural model is used for modelling monthly average streamflow series of a Brazilian river. Experimental results show good performance of the proposed approach.
Partial mutual information; input selection; Kernel functions; non-parametric regression; temporal series
