Sparse Graphical Vector Autoregression: A Bayesian Approach

What is it about?

This paper proposes a model selection approach to multivariate time series of large dimension, by combining graph-based notion of causality with the concept of sparsity on the structure of dependence among the variables. We build on the application of fan-in restriction for graphical models by proposing a sparsity-inducing prior that allows for different information level about the maximal number of predictors for each equation of a VAR model.

Featured Image

Why is it important?

We demonstrate the effectiveness of our approach through simulation experiments and empirical applications to macroeconomics and finance. We compare the results from our model with the one obtained with standard Lasso-type methods. We find evidence that our new prior distribution induces sparsity on the temporal dependence between variables and produces networks that are more parsimonious and presents a better representation of the temporal dependence between variables than the Lasso-type methods. We show that the MSFE are substantially equal across the methods, whilst the density forecasts favor our sparse graphical model over the Lasso-type ones.