Graphical spike-and-slab priors for network-based VAR regularization
||Graphical spike-and-slab priors for network-based VAR regularization|
||SNIC Small Compute|
||Héctor Rodriguez Déniz <firstname.lastname@example.org>|
||2022-09-12 – 2023-02-01|
Modeling problems using a network representation has become common practice in science to deal with the complexity of many real systems, such as the internet, social media, and public transportation. The increasing availability of data has not only contributed to widespread the research and use of network-based methodologies, but also to pose new challenges such as the development of effective methods to model and forecast dynamic network processes by using time series of graph data. The Vector Autoregression (VAR) is probably the most popular linear model for multivariate time series data and has been thoroughly studied and applied in many research areas. However, VAR models suffer from overparameterization when the number of output functions grows, which is inevitable when working with graphs. In this project we propose a graphical spike-and-slab regularization strategy for the Bayesian VAR by assuming functional dependencies that can be represented as directed graphs. Both the VAR and the latent network-based regularization model are estimated jointly via Gibbs sampling. We perform simulation tests on synthetic data, and present two real-world applications using macroeconomic time series data.