Amisano-Federico (2004) Alternative Time-Varying Parameter Specifications for Bayesian VAR Models

Summary

Amisano and Federico (2004) compare four specifications for the state equation error covariance $\Omega$ in Bayesian vector autoregression (VAR) models with time-varying parameters (VAR-TVP), applied to Euro area inflation forecasting. The central finding is that parsimony in specifying $\Omega$ is crucial: the Kronecker-structure model ($\Omega = \rho(R \otimes Q_k)$), which ties coefficient dynamics to the observation covariance, yields the best out-of-sample performance while also simplifying the Kalman filter. The fully general Cogley-Sargent specification (cross-correlated observation and state errors) is the worst model. Marginal likelihood computation methods are compared and found to break down for high-dimensional parameter spaces.

Key Claims

• In a VAR-TVP model the state equation error covariance $\Omega$ governs the degree and cross-equation structure of coefficient time variation; more parsimonious $\Omega$ specifications outperform richer ones in out-of-sample forecasting.
• The Kronecker structure $\Omega = \rho(R \otimes Q_k)$ (M3 model) achieves Theil's-U gains of 9% at $h = 2$ and 12% at $h = 4$ over a fixed-coefficient Bayesian VAR (BVAR) for Euro area inflation; it also simplifies the multimove Gibbs sampler by reducing the $(K \times K)$ Kalman matrix inversion to a $(k \times k)$ inversion at each step.
• The fully general M1 model (Cogley-Sargent style, with $C \neq 0$ cross-correlation between observation and state errors) is the worst performer — its prior is difficult to calibrate and it ranks last or near-last on all accuracy measures.
• Estimating the time-variation hyperparameter $\rho$ rather than fixing it at a default value (M4 vs. M4,0) improves long-horizon inflation forecasting.
• Marginal likelihood (ML) computation: Chib's (1995) method and the Laplace method agree closely for low-dimensional models and are both accurate; Gelfand-Dey (1994) modified importance sampling is more numerically stable but slightly less precise. All methods break down for M3 (Chib returns an implausibly large ML), confirming that no fully satisfactory ML estimator exists at high parameter dimension.
• The Savage-Dickey density ratio provides an exact Bayes factor for nested comparisons (e.g., M4,0 vs. M4, testing $h = h_0$) from the marginal posterior of the restricted parameter alone.

Concepts Introduced or Extended

• Time-Varying Parameter VAR — state-space formulation; four $\Omega$ specifications; Kronecker structure; Carter-Kohn multimove sampler; Kalman filter equations
• Vector Autoregression — Bayesian VAR base model with time-varying coefficients
• Gibbs Sampler — Carter-Kohn multimove algorithm for joint state sampling; Kronecker Kalman filter simplification
• Marginal Data Density — Gelfand-Dey modified importance sampling; Laplace method; Savage-Dickey density ratio; breakdown at high dimension

Entities Mentioned

• Gianni Amisano
• Loredana Federico

Quotes

"we have found satisfactory performances in models that are based on sensible hypotheses that greatly reduce the number of free parameters in the state equation."

"Two such methods are a development of Litterman's original approach, based on the use of a very small subset of parameters, and the imposition of a Kronecker structure on the variance-covariance matrix of state equation errors."

My Take

The Kronecker structure (M3) is the paper's practical contribution: it enforces a reasonable prior belief (coefficient dynamics scale with observable variance) and as a side effect collapses the computational bottleneck in the Kalman filter. The finding that M1 is the worst model is a strong practical warning against over-parameterizing $\Omega$ in TVP-VARs. The marginal likelihood section doubles as a useful applied comparison of three numerical estimators, though the failure at large $p$ means it cannot resolve the most interesting model comparisons (those involving M3). The paper predates Primiceri (2005) and Cogley-Sargent (2005), which became the standard TVP-VAR references, but establishes much of the same Gibbs machinery.