Selecting GARCH Volatility Models by Intended Use

Summary

The document asks whether volatility models with statistically significant parameter estimates should be preferred over otherwise similar specifications whose estimates are insignificant. The examples include sGARCH, EGARCH, and TGARCH models fitted with alternative innovation distributions, such as skewed normal and Student-t variants. The questioner proposes retaining the specifications that produce significant alpha estimates and excluding those that do not.

The response says model choice depends on the purpose of the analysis. If the goal is forecasting, significance alone is not an adequate selection rule; model fit and predictive performance on holdout data should also be considered. This shifts attention from a single coefficient test to evaluation aligned with the intended use. The answer does not prescribe specific metrics, validation procedures, or a preferred model, and the exchange provides no empirical comparison of the listed specifications. Its central guidance is therefore a model-selection principle rather than evidence that one volatility model or innovation distribution is superior.

Key ideas

• The appropriate volatility model depends on the intended use of the analysis.
• A significant parameter estimate alone does not establish that a specification is the best choice.
• For forecasting, compare model fit and predictive accuracy on holdout samples.
• The discussion does not identify a universally preferred GARCH form or innovation distribution.

Tags

# Volatility models using Rugarch


# Volatility models using Rugarch












I have estimated sGARCH, EGARCH and TGARCH, which some for particular models are significant. For others, the alpha remain insignificant using various innovations such as the skewed variants of the normal or student $t$ distributions.

I am tempted to rely on the models that give significant estimates and then ignore the ones that do not. For instance, if the EGARCH-std is not significant in terms of alpha and the EGARCH-sstd is significant, I prefer to focus on the later for further analysis.

Please let me know if I am on the right path.

## Answer by Ryogi (score 1)

https://quant.stackexchange.com/a/4436

The answer depends on what you will use the models for. For example, if you care about prediction, you should use different metrics (model fit, prediction accuracy on hold-out samples, etc) to determine what works best.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.