Bayesian TGARCH (Threshold GARCH with Bayesian Estimation)
Bayesian Threshold Generalized Autoregressive Conditional Heteroscedasticity Model · Also known as: Bayesian TGARCH, Bayesian GJR-GARCH, Threshold GARCH with Bayesian estimation, TGARCH-B
Bayesian TGARCH combines the Threshold GARCH volatility model — which captures the asymmetric response of volatility to positive versus negative shocks — with full Bayesian inference via Markov Chain Monte Carlo sampling. The result is a principled, uncertainty-aware framework for modeling leverage effects and fat-tailed financial returns.
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 Bayesian TGARCH when modeling financial return series where you expect asymmetric volatility (leverage effects), and you want full posterior uncertainty quantification rather than point estimates. It is particularly suited to small or moderate samples where MLE can produce boundary solutions or imprecise standard errors, and when prior information from similar markets or assets is available. It is also appropriate when you need proper interval estimates for risk measures such as Value-at-Risk or Expected Shortfall. Avoid it when data are very abundant and computational speed is a priority, or when there is no substantive reason to expect asymmetric volatility — in those cases a standard Bayesian GARCH is simpler and equally valid.
Strengths & limitations
- Captures the leverage effect: negative shocks raise future volatility more than positive shocks of equal size.
- Full posterior distributions provide honest uncertainty quantification for all parameters and derived risk measures.
- Priors can encode stationarity constraints, preventing invalid parameter estimates that MLE sometimes produces.
- Works well with small samples where maximum-likelihood TGARCH is unreliable.
- Easily extended to Student-t or other heavy-tailed error distributions within the same MCMC framework.
- Posterior predictive distributions directly yield probabilistic forecasts of future volatility.
- Computationally expensive: MCMC sampling is orders of magnitude slower than MLE for large datasets or long chains.
- Results are sensitive to prior choice, especially with limited data; poorly specified priors can dominate inference.
- Convergence diagnostics (trace plots, Gelman-Rubin R-hat) must be checked carefully; failures are not always obvious.
- Implementation requires specialized software (e.g., bayesGARCH in R or custom Stan/PyMC code).
- Interpreting and communicating posterior distributions is more demanding than reporting point estimates with standard errors.
Frequently asked
What is the threshold in TGARCH and why does it matter?
The threshold is the value of zero for the lagged shock: the indicator 1_{eps<0} switches on an extra volatility term only when the previous period's shock was negative. This captures the leverage effect — the well-documented tendency for bad news to raise volatility more than equally sized good news.
How does Bayesian TGARCH differ from maximum-likelihood TGARCH?
MLE produces a single point estimate for each parameter and asymptotic standard errors. Bayesian estimation produces full posterior distributions, yielding credible intervals that reflect genuine parameter uncertainty, are valid in small samples, and can incorporate prior information. The trade-off is substantially higher computation time.
What prior should I use for the asymmetry parameter gamma?
A common choice is a Normal or truncated Normal prior centered at zero with moderate variance, expressing mild prior belief that the leverage effect may exist but is not assumed to be large. If you have strong evidence from comparable assets, an informative prior can improve estimation in small samples.
How do I know if my MCMC chain has converged?
Inspect trace plots for all parameters (they should look like 'hairy caterpillars' with no trending), compute the Gelman-Rubin R-hat statistic (values below 1.05 are acceptable), and check that the effective sample size (ESS) is large enough — at least 400 is a common rule of thumb for reliable posterior summaries.
Can Bayesian TGARCH handle non-normal innovations?
Yes. A Student-t or Generalized Error Distribution for the innovations is easily incorporated by placing a prior on the degrees-of-freedom or shape parameter and sampling it jointly with the variance parameters. This is one of the key practical advantages of the Bayesian approach over closed-form MLE derivations.
Sources
- Zakoian, J.-M. (1994). Threshold heteroskedastic models. Journal of Economic Dynamics and Control, 18(5), 931-955. DOI: 10.1016/0165-1889(94)90039-6 ↗
- Ardia, D. (2008). Financial Risk Management with Bayesian Estimation of GARCH Models: Theory and Applications. Springer. ISBN: 978-3-540-78656-6
How to cite this page
ScholarGate. (2026, June 3). Bayesian Threshold Generalized Autoregressive Conditional Heteroscedasticity Model. ScholarGate. https://scholargate.app/en/econometrics/bayesian-tgarch
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.
- Bayesian ARCH modelEconometrics↔ compare
- Bayesian EGARCHEconometrics↔ compare
- Bayesian GARCH modelEconometrics↔ compare
- DCC-GARCH modelEconometrics↔ compare
- EGARCH modelEconometrics↔ compare
- TGARCH modelEconometrics↔ compare