I have the following question: I am analyzing Brent Oil returns and I have found that there's a significant negative sign bias. So first, I have tried with a GARCH(1,1) and it turns out that completly cleans the ACF, so there are no significant autocorrelations left, and adjusts quite decently to the data.
However, due to the presence of negative sign bias I've decided to perform a TARCH(1,1) but it turns out that performs dramatically worse than GARCH(1,1), as ACF is not cleant and residuals still have lot of noise.
Any suggestion of how I should tackle this issue?
Thank you very much!
Best Answer
Citing the vignette of "rugarch" package in R,
If one, several or all the hypotheses are rejected, the idea is to use a model that allows for asymmetric effects such as GJR-GARCH, APARCH, TARCH. (Hence, you may try the ones you have not tried yet.)
However, if you are unable to obtain a satisfactory model that allows for asymmetry, perhaps using a relatively good model without assymetry will be a lesser evil than using a poor asymmetric model. The model choice is essentially an empirical question. You could try doing out-of-sample performance evaluation to select the model that suits your needs best.
References