Do we go down the route of putting yield curves expect to see in the dataset ?
One concern here is that we transform the dataset to model the change in yield over a 1 year time step.
So to include a year of parallel downward shift say, you could add one years movement.
But this inclusion feels arbitrary in nature since its weight is dictated by the existing size of the calibration dataset. And the whole purpose of the bayesian framework is to be more quantitative and less arbitrary.