This is still on our todo-list. They issue is that you would need to draw from the posterior distribution of parameters and then iterate forward on the VAR as you suggested. Your approach would neglect parameter uncertainty, which can be sizeable. A first approach by a user was at Writtng a dsgevar_forecast function