Real-time conditional forecasts with Bayesian VARs: An application to New Zealand

We develop a large Bayesian VAR (BVAR) model of the New Zealand economy that incorporates the conditional forecasting estimation techniques of Waggoner and Zha (1999). We examine the real-time forecasting performance as the size of the model increases using an unbalanced data panel. In a realtime out-of-sample forecasting exercise, we find that our BVAR methodology outperforms univariate and VAR benchmarks, and produces comparable forecast accuracy to the judgementally-adjusted forecasts produced internally at the Reserve Bank of New Zealand. We analyse forecast performance and find that, while there are trade offs across different variables, a 35 variable BVAR generally performs better than 8, 13, or 50 variable specifications for our dataset. Finally, we demonstrate techniques for imposing judgement and for forming a semi-structural interpretation of the BVAR forecasts.