Time-varying parameter estimation in macro-econometrics

The paper addresses the problem and related issues of Time-Varying Parameter (TVP) estimation, a technique recently introduced in the field of Macro-Econometrics, and especially in FAVAR (Factor-Analysis Vector Auto-Regression) modeling. Different from standard multiple or single regression estimation, where Time-Fixed Parameter (TFP) estimation dominates over the entire sample and may be conducive to the “Lucas Critique”, TVP produces changing parameters which may be utilized by the analyst to infer the dynamics underlying the data process, such as structural breaks, changes in covariances, parameter significance, and so on. This advantage, however, comes at a high cost represented by two major occurrences. The first is the degree of attentiteveness placed by the analyst in constructing the formal building blocks of FAVAR models, while the second is the machine timing required to produce Gibbs-simulated Normal and inverse Wishart parameter distributions to approximate Bayesian priors and posteriors utilized in the underlying Kalman processes. Sizable costs may also ensue from the treatment of synchronous parameters and from the construction of impulse responses and variance decompositions for the purpose of policy interpretation. In- and out-sample forecasting applied to competing TVP and TFP FAVAR models of the US economy and monetary policy during the years 1953-2011, using quarterly observations, produces very interesting results that tend to favor the TVP approach.