This work deals with an analysis of Granger Causality computation based on artificial neural networks, including a nonlinear relation between the involved variables. Information about the training parameters are exhibited in order to visualize how the conditions of the chosen model to obtain the connectivity information depend on the architecture of network. Three chaotic maps with a bivariate case built from two time series were employed to see the effect of training parameters of the models. Nonlinear autoregressive and nonlinear autoregressive with exogenous inputs were used to forecast the time series, and then, obtain the causality information based on differences of errors between both approximations. Results show that the causality computation is sensible to neural network parameters previously untreated in a detailed mode
This work deals with an analysis of Granger Causality computation based on artificial neural networks, including a nonlinear relation between the involved variables. Information about the training parameters are exhibited in order to visualize how the conditions of the chosen model to obtain the connectivity information depend on the architecture of network. Three chaotic maps with a bivariate case built from two time series were employed to see the effect of training parameters of the models. Nonlinear autoregressive and nonlinear autoregressive with exogenous inputs were used to forecast the time series, and then, obtain the causality information based on differences of errors between both approximations. Results show that the causality computation is sensible to neural network parameters previously untreated in a detailed mode.
