Predicting Time-Varying Parameters with Parameter-Driven and Observation-Driven Models
37 Pages Posted: 6 Mar 2012
Date Written: March 5, 2012
We study whether and when parameter-driven time-varying parameter models lead to forecasting gains over observation-driven models. We consider dynamic count, intensity, duration, volatility and copula models, including new specifications that have not been studied earlier in the literature. In an extensive Monte Carlo study, we find that observation-driven generalised autoregressive score (GAS) models have similar predictive accuracy to correctly specified parameter-driven models. In most cases, differences in mean squared errors are smaller than 1% and model confidence sets have low power when comparing these two alternatives. We also find that GAS models outperform many familiar observation-driven models in terms of forecasting accuracy. The results point to a class of observation-driven models with comparable forecasting ability to parameter-driven models, but lower computational complexity.
Keywords: Generalised autoregressive score model, Importance sampling, Model confidence set, Nonlinear state space model, Weibull-gamma mixture
JEL Classification: C53, C58, C22
Suggested Citation: Suggested Citation