Bayesian Bi-level Sparse Group Regressions for Macroeconomics Forecasting
Economics
Further details
This paper considers forecasting models for macroeconomic aggregates in a data rich
environment where predictors are organized in groups. Variables in each group are characterized
by strong covariation. Due to the large dimension, some groups and some predictors
in a group might be not relevant to forecast the target variable, conditional on the
remaining groups and predictors in the same group. In this paper we take advantage of this
group structure by constructing a hierarchical prior distribution that treats the coefficients
of each block independently, imposes correlation among the coefficients in each block
and induces bi-level sparsity. We show that this prior is computationally convenient. We
demonstrate the theoretical validity of our Bayesian procedure from a frequentist point
of view and show that it attains the optimal rate of convergence. The setting is extended
further to take into account stochastic volatility. Finally, we illustrate finite sample properties
of our procedure through Monte Carlo experiments and analyse a real data forecasting
problem.
