Skip to content

Identifiability in GAS models

Home / Open Questions / Identifiability in GAS models

Background: Generalized Autoregressive Score (GAS) models for time-varying transition probabilities are defined by scaling mechanisms that can couple regime-specific variances with transition probability parameters. This coupling often creates non-identifiability in the likelihood surface.

Question / Future Work: The identification of score coefficients in GAS-based regime-switching models remains challenging due to the emergence of ridges in the joint likelihood surface of regime variances and transition dynamics, necessitating further research into robust estimation techniques or alternative parameterizations.

Why It Matters: Understanding this identifiability issue is crucial for practitioners attempting to use GAS-driven Markov-switching models, as it explains why these models frequently fail to converge or collapse to simpler constant-transition specifications in empirical applications.

Evidence: The GAS score coefficient appears to be statistically non-identifiable, due to a ridge in the joint likelihood surface (σ^2, A).

Metadata & Links