Abstract

In this paper, we combine density forecasts from Bayesian quantile regressions. We develop a forecasts combination scheme that assigns weights to the individual predictive density forecasts based on quantile scores. Compared to standard combination schemes, our approach has the advantage of assigning different set of combination weights to the various quantiles of the predictive distribution. We apply our approach to US GDP growth forecasts based on quantile regressions using a broad set of common leading indicators. The results show that density forecasts from our quantile combination approach outperforms forecasts from commonly used combination approaches such as Bayesian Model Averaging, optimal combination, combinations based on recursive logarithmic score weights and equal weights. In particular, our quantile combination approach provides more accurate forecasts for the lower tail of the GDP distribution, measuring downside macroeconomic risk.