Comparing Impulse Response Estimation across VAR and Local Projection Methods
Аннотация
This paper examines impulse response estimation in macroeconomic analysis using Vector Autoregression (VAR), Local Projections (LP), Empirical-Bayes Minnesota BVAR (EB-BVAR), Penalised Local Projections (PLP), and Bayesian Network Vector Autoregression (BNVAR). We conduct Monte Carlo simulations calibrated to a U.S. macroeconomic dataset and consider sparse, dense, moving-average, and network data-generating processes. The results reveal substantial differences in finite-sample performance across estimators and highlight the importance of regularisation. BNVAR performs particularly well under sparse and network structures, while EB-BVAR achieves the lowest RMSE under the dense and moving-average designs. PLP consistently improves upon standard LP. Overall, the findings show that the relative performance of impulse response estimators is strongly dependent on the underlying dynamic structure, with BNVAR providing a competitive regularised approach, particularly in sparse and network-driven environments.