This research proposes a novel method for estimating Impulse Responses (IRFs) by doing Bayesian inference directly on the parameters of Vector Moving Average (VMA) representation of the large-scale VAR. Small scale VARs are plagued with omitted variable bias in residuals, hence the use of models with many variables, profiting from big data revolution. On the flip-side, estimating sheer number of model parameters and IRF coefficients is a daunting task. We aim at developing an automatic algorithm for eliciting priors on the IRF coefficients directly, improving the precision of estimates. Shrinkage of the parameter space is done using stochastic search variable selection (SSVS), a hierarchical prior dependent on a single hyperparameter, which is to be estimated from data. This allows for higher transparency and more objective shrinkage in the Bayesian context. Algorithm is to be tested on estimating responses from monetary policy disturbance in case of a foresight of policy actions by economic agents.
Constraints econometricians and decision makers face in estimation and in modeling the economy led them to take a look at new methods and practices. Profiting from flexibility of VARs, exploiting the availability of big data resources and introducing new exciting concepts of pattern recognition in data and machine learning is a way to look for in the future (as discussed in Varian, 2014 and Ramey, 2016 among many others). This is a new, open and dynamic ecosystem for conducting economic analysis, with exciting new methods, which are already reforming the way we do quantitative macroeconomic modeling.
This research brings together above mentioned concepts and explores the possibility to improve the precision of our estimates on the effects of policy measures, can contribute to better understanding of shock propagation mechanism and might be readily applied in practice.
Our main contribution is a mechanism of eliciting more objective and transparent prior with respect to previous practice (e.g. Minnesota approach). Moreover, automation of the process should improve the precision, especially in models of medium and large scale. Instead of imposing priors on the VAR parameters, which indirectly affect and distort impulse responses, we do the inference directly on the coefficients of VMA, i.e. impulse responses themselves. Push to include all relevant information in a VAR framework, either through FAVARs or Large-scale Bayesian VARs led to the best models on average that we have when it comes to forecasting performance. This is a strand of research exercising fast development in the last couple of years, with promising prospects in the future.