A1846
Title: Granger-causality detection in high-dimensional systems using feedforward neural networks
Authors: Jose Olmo - Universidad de Zaragoza (Spain) [presenting]
Hector Calvo-Pardo - University of Southampton (United Kingdom)
Tullio Mancini - University of Southampton (United Kingdom)
Abstract: A novel methodology is proposed to detect Granger causality in mean using feedforward neural networks. The approach accommodates unknown dependence structures between the elements of highly-dimensional multivariate time series with weak and strong persistence characterized by a vector of lags potentially increasing to infinity. To do this, we propose a two-stage procedure. In a first stage, we fit a neural network given by an optimal number of nodes in the intermediate hidden layers. This is done by minimizing the entropy of the neural network - maximizing the transfer of information between input and output variables. In a second stage, we apply a novel sparse double group lasso penalty function to identify the variables that have predictive ability in the multivariate time series and, hence, Granger cause of the others. The penalty function inducing sparsity is applied to the weights characterising the nodes of the neural network and allows us to add interpretability to the neural network by mapping the nodes of the neural network to the variables exhibiting Granger causality. We show the correct identification of these weights when the number of variables and lags is finite and also when increases to infinity with the sample size. An application to the recently created Tobalaba network of renewable energy companies shows how to exploit Granger-causality for uncovering the role of the net in increasing connectivity between companies and forecast ability.
