Modelling the Relation between the US Real Economy and the Corporate Bond-Yield Spread in Bayesian VARs with non-Gaussian Disturbances

In this paper we analyze how skewness and heavy tails a ect the estimated relationship between the real economy and the corporate bond-yield spread, a popular predictor of real activity. We use quarterly US data to estimate Bayesian VAR models with stochastic volatility and various distributional assumptions regarding the disturbances. In-sample, we find that after controlling for stochastic volatility innovations in GDP growth can be well-described by a Gaussian distribution. In contrast, both the unemployment rate and the yield spread appear to benefit from being modelled using non-Gaussian innovations. When it comes to real-time forecasting performance, we find that the yield spread is an important predictor of GDP growth, and that accounting for stochastic volatility matters, mainly for density forecasts. Incremental improvements from non-Gaussian innovations are limited to forecasts of the unemployment rate. Our results suggest that stochastic volatility is of first order importance when modelling the relationship between yield spread and real variables; allowing for non-Gaussian innovations is less important.