Using innovative methods: Crises like the late 2000s financial crisis or, most recently, the Covid-19 pandemic, stressed the need to better understand the complex dynamics of the global economic and financial system. “Existing empirical methods only allow us to do this to a limited extent, however. As a consequence, the complex patterns of how monetary policy decisions affect the economy cannot be fully captured,” explains Gregor Kastner, who submitted the project jointly with Luis Gruber (both from the Department of Statistics at the University of Klagenfurt). The result: relevant financial and economic linkages which play a fundamental role in the spread of systemic risk are often overlooked.

To create a better overall picture in the future, better methods are needed to enable researchers to fully extract the relevant information required for forecasting from the ever-increasing amount of data. Gregor Kastner goes on: “In our project, we propose new approaches to address these challenges, with the concept of structured covariance modeling as the common denominator.”

From a methodological perspective, the research team delivers three fundamental contributions: “First, we provide a profound analysis of state of the art Bayesian multivariate stochastic volatility specifications and demonstrate that their forecasting performance can substantially be improved by minor adaptions in the form of structured prior distributions. Secondly, we focus on Bayesian factor stochastic volatility models and the notoriously hard question of dynamic factor selection. Third, we propose two novel Bayesian dynamic graphical models which, other than traditional covariance models, model the inverse of the covariance, the precision.”

As they explain, Gregor Kastner and Luis Gruber will next apply the proposed methodology to complex financial and macroeconomic data sets: “On the one hand, we focus on probabilistic forecasting to perform model validation. On the other hand, we demonstrate that our methods can be used to answer relevant questions of central banks and other policy institutions.”

In the project and in the context of dynamic covariance modeling, the research team aims to integrate “classical” (“foundational”) ideas from the early stages of Bayesian modeling, such as the inclusion of structured domain knowledge in the form of prior distributions, and ideas of automated learning. “The synthesis of these worlds is what we label semi-automatic learning”, as Gregor Kastner explains.