Do so-called multivariate filters have better revision properties? An empirical analysis

The output gap plays a crucial role in thinking and actions of many central banks but real time measurements undergo substantial revisions as more data become available (Orphanides (2001), Orphanides and van Norden (forthcoming)). Some central banks augment, such as the Bank of Canada and the Reserve Bank of New Zealand, the Hodrick and Prescott (1997) filter with conditioning structural information to mitigate the impact of revisions to the output gap estimates. In this paper, we use a state space Kalman filter framework to examine whether the augmented (so-called “multivariate filtersâ€) achieve this objective. We find that the multivariate filters are no better than the Hodrick-Prescott filter for real-time NZ data. The addition of structural equations increase the number of signal equations, but at the same time adds more unobserved trend/equilibrium variables to the system. We find that how these additional trends/equilibrium values are treated matters a lot, and they increase the uncertainty around the estimates. In addition, the revisions from these models can be as large as a univariate Hodrick-Prescott filter.