I am writing my master's thesis and using a AR(1) GJR-GARCH(1,1)-EVT-Copula model on my data. One of the main papers I use is McNeil & Frey (2000), in which they only do AR-GARCH-EVT. In this paper, for the GARCH part, McNeil & Frey insist that they use Pseudo Maximum Likehood (or Quasi Max Likelihood) to estimate paramaters to avoid too many assumptions on the distribution of $z_q$.

However, in other papers, people use MLE to estimate GJR-GARCH model and then go to the copula part. Additionally, Matlab does MLE estimation very easily via the estimate command so it would be a huge time saver.

My question is: Is QML really a much better method for estimating the GJR-GARCH parameters or is MLE working just fine in this case?

A clarification would be much appreciated :)



Hi and welcome to Cross Validated!

The difference between the MLE and QML is rather subtle, but it has to do with the assumption of normality and its appropriateness. In the GARCH setting, the assumption of normality of returns is particularly useful because it significantly simplifies calculations in the likelihood function. However, it is also a "well-established" fact that financial returns are rarely normally distributed, even over longer time horizons when one would expect the Central Limit Theorem effects to kick in.

To quote McNeil, Frey & Embrechts themselves, from their book Quantitative Risk Management - Concepts, Techniques and Tools, Revised edition, on section 4.2.4 Fitting GARCH Models to Data, p. 146:

In describing the behaviour of parameter estimates in the following paragraphs, we distinguish two situations. In the first situation we assume that the model that has been fitted has been correctly specified, so that the data are truly generated by a time-series model with both the assumed dynamic form and innovation distribution. We describe the asymptotic behaviour of the maximum likelihood estimates (MLEs) under this idealization.

In the second situation we assume that the correct dynamic form is fitted but that the innovations are erroneously assumed to be Gaussian. Under this misspecification, the model fitting procedure is known as quasi-maximum likelihood (QML) and the estimates obtained are QMLEs. Essentially, the Gaussian likelihood is treated as an objective function to be maximized rather than a proper likelihood; our intuition suggests that this may still give reasonable parameter estimates, and this turns out to be the case under appropriate assumptions about the true innovation distribution.

In other words, you can use the typical ML when you know that your $z_q$ do indeed follow a normal distribution. Otherwise, when you can see, for example via appropriate tests, that your $z_q$ are not normally distributed, you should use QML (as long as the innovations are iid with zero mean and unit variance).

  • $\begingroup$ Ok thank you for the clarification. Then yes I will probably use QML. I think from the documentation that actually tarch command from the MFE toolbox does etimate a GJR-GARCH by QML. If anyone knows for sure, I'd love to hear a confirmation ^^ $\endgroup$ – Edge284 Jul 11 '19 at 13:31
  • $\begingroup$ If the answer is satisfactory and covers your question, please mark it as accepted so that it can be closed. $\endgroup$ – Emil Jul 11 '19 at 13:44

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.