will give a presentation on
VARs have had great success in macroeconomic forecasting. But most of the existing literature evaluates their forecast performance based either on point forecasts or entire predictive densities. Recently, there has been a growth of interest in forecasting the tails of predictive densities so as to produce measures of tail risk to economic outcomes. Linear VARs with Normal homoscedastic errors are typically found not to produce accurate measures of tail risk. In this paper, we compare various parametric and non-parametric extensions of VARs in their ability to accurately forecast tail risk. The parametric extension involves stochastic volatility. The non-parametric extensions involve regression trees. We develop regression tree specifications for both the conditional mean and conditional variance and investigate the relative importance of each in forecasting tail risk.
This is joint work with Todd Clark, Florian Huber, Michael Pfarrhofer and Massimiliano Marcellino.