Dimensionality reduction in macroeconometrics models

Standard vector autoregressive approach is highly used in the macroeconometrics. However, the problem with degrees of freedom arises. It is usually quite challenging to include a lot of information in the model. Subsequently, the predictive power declines, and the structural analysis may be misleading. In general, even though there is plenty of available macroeconomic series, we can operate with only an insufficient portion. More generally, the thesis contains three advanced methodology guidelines included Bayesian, machine learning, and econometric approaches. These approaches deal with dimensional reduction according to the two different methodology procedures. The first procedure reduces the predictors' space with a few common unobserved factors. On the contrary, the second one is more focused on shrink the parameter space based on the particular criteria. In the thesis, we empirically examine the Czech main macroeconomic indicators with a main focus on the unemployment rate. We found that the advanced approaches produce effective unemployment rate forecasts compared to the standard ARMA or VAR models. Moreover, we capture the individual impacts of the indicators on the unemployment rate using the structural VAR analysis.