Time-Varying Parameter GARCH Model (TVP-GARCH)
Time-Varying Parameter Generalized Autoregressive Conditional Heteroscedasticity Model · Also known as: TVP-GARCH, time-varying GARCH, TV-GARCH, state-space GARCH
The Time-Varying Parameter GARCH model extends the standard GARCH framework by allowing the conditional variance parameters — including the ARCH and GARCH coefficients — to change over time rather than remaining fixed throughout the sample. This makes it well-suited to financial and macroeconomic series where volatility dynamics evolve across different market regimes or economic episodes.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use TVP-GARCH when you have a long financial or macroeconomic time series and have evidence — from structural break tests, rolling-window estimates, or economic reasoning — that volatility dynamics are not stable over time. It is particularly valuable around financial crises, policy regime changes, or other episodes where risk characteristics shift substantially. Do not use it when the sample is short (fewer than ~200 observations), as the filter cannot reliably track parameters with limited data. Avoid it when a simpler GARCH specification with dummy variables for known break dates is adequate, or when out-of-sample forecasting is the primary goal and parsimony matters more than flexibility.
Strengths & limitations
- Captures regime changes and structural breaks in volatility dynamics without requiring knowledge of break dates.
- More realistic than fixed-parameter GARCH for long financial samples spanning multiple business and credit cycles.
- The GAS/score-driven variant is computationally tractable and does not require a full state-space recursion.
- Provides time-stamped coefficient estimates that can be interpreted economically alongside market events.
- Nests the standard GARCH model as a special case when parameter variance Q equals zero.
- Requires long samples (at least 200–300 observations) for the filter to reliably identify parameter drift from noise.
- Significantly more parameters and hyperparameters to estimate than standard GARCH; optimisation can be numerically fragile.
- Identified parameter paths depend heavily on the assumed transition equation (random walk vs. mean-reverting), which is often difficult to test.
- Out-of-sample forecasting is more complex than with fixed-parameter models because future parameter values must also be predicted.
- Interpretation is challenging when multiple parameters vary simultaneously, making it hard to separate the sources of changing volatility.
Frequently asked
How does TVP-GARCH differ from Markov-switching GARCH?
Markov-switching GARCH assumes the parameters jump between a small number of discrete regimes according to a hidden Markov chain. TVP-GARCH instead allows parameters to drift continuously and smoothly. Use Markov-switching when you believe volatility shifts abruptly between well-defined states; use TVP-GARCH when you expect a gradual, ongoing evolution of volatility dynamics.
What is the GAS (score-driven) approach and how does it relate to TVP-GARCH?
The Generalized Autoregressive Score (GAS) framework of Creal, Koopman, and Lucas (2013) updates time-varying parameters using the scaled score of the predictive log-likelihood. Applied to GARCH, this produces a TVP-GARCH variant that is robust, computationally simple, and does not require a Kalman filter, making it a popular alternative to the full state-space formulation.
How many observations do I need for reliable TVP-GARCH estimation?
As a practical guideline, at least 200–300 observations are needed for the Kalman filter to separate genuine parameter drift from estimation noise. With shorter samples the estimated paths are dominated by filter uncertainty, and simpler fixed-parameter GARCH models typically outperform TVP-GARCH both in-sample and out-of-sample.
Can TVP-GARCH be extended to multivariate series?
Yes. Multivariate extensions allow the conditional covariance matrix to have time-varying ARCH and GARCH loadings, which is useful for dynamic conditional correlation modelling. However, the number of parameters grows rapidly with the number of series, so careful regularisation or factor-based dimensionality reduction is usually needed.
How do I choose between a random-walk and a mean-reverting transition equation for the parameters?
A random walk is the most common default because it is parsimonious and requires no additional hyperparameters. A mean-reverting (Ornstein-Uhlenbeck) specification is preferable when you have economic reasons to believe parameters fluctuate around a long-run equilibrium value. You can compare the two specifications using the Akaike or Bayesian information criterion.
Sources
- Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation. Econometrica, 50(4), 987-1007. DOI: 10.2307/1912773 ↗
- Creal, D., Koopman, S. J., & Lucas, A. (2013). Generalized autoregressive score models with applications. Journal of Applied Econometrics, 28(5), 777-795. DOI: 10.1002/jae.1279 ↗
How to cite this page
ScholarGate. (2026, June 3). Time-Varying Parameter Generalized Autoregressive Conditional Heteroscedasticity Model. ScholarGate. https://scholargate.app/en/econometrics/time-varying-parameter-garch-model
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- EGARCH modelEconometrics↔ compare
- GARCH ModelEconometrics↔ compare
- Kalman FilterBayesian↔ compare
- State Space ModelEconometrics↔ compare
- Stochastic Volatility ModelFinance↔ compare