Title: Large Bayesian VAR forecasting in a national accounting environment
Authors: Alexander Rathke - ETH Zurich, KOF Swiss Economic Institute (Switzerland) [presenting]
Samad Sarferaz - ETH Zurich (Switzerland)
Abstract: Large Bayesian VARs are used, including national accounting identities from the expenditure side and the production side, i.e. the demand and supply side. Hence, we produce forecasts for GDP that fully rely on the data generating process. We show that by using national accounting identities, we outperform smaller VARs that only use a final GDP measure, instead of the aggregated one. We further show how to implement expert based priors on the expenditure and production side using soft and hard conditions. For the expenditure side we use expert based knowledge and for the production side we use business tendency surveys conducted in different sectors to inform our priors. Given that the forecasting exercise results in two different forecasts for GDP, we further combine these two estimates using different weighting schemes, which can differ across different frequencies and horizons. We show that a combination of these two estimates is beneficial. Interestingly, the optimal weight for the combination of the two GDP measures gives the expenditure side more in the short-run and the production side more weight the medium-term.