Has the Covid-19 outbreak capsized the predictive performance of Bayesian VAR models with cointegration and time-varying volatility?

We check whether taking into account long-term relationships in heteroscedastic VAR models affects their predictive performance before and during the Covid-19 pandemic. Also, we examine whether the predictions can benefit from suspending posterior updates at some point. Empirical analysis covers five different economies and uses Bayesian VAR/VEC models with volatility specifications combining stochastic volatility and GARCH processes. It emerges that, while accounting for cointegration relationships in the models enhances their predictive performance prior to the pandemic, it may be counterproductive for times of economic crisis. Additionally, refraining from keeping the posterior updated does improve the predictions, but only rarely.