Why can't I remove the ARCH effect and Serial Correlation?

[chauhan7]

Hi

Please see the attachment.

The file contains NASDAQ 100 stock market returns, where there are dummy variables to indicate each day of the week from monday to friday, where the objective is to find anomalous behaviour, i.e. days which exhibit abnormal returns. However, unlike my UK data, the US data I have appears to show an immense amount of heteroskedasticity and autocorrelation, both of which I cannot remove as it appears adding ar(1) has little or no effect. I'm sure the GARCH model is supposed to remove all heteroskedasticity anyway, I just have no idea what to do with this data now.

Thanks

[startz]

What is it the makes you think you have an immense amount of autocorrelation?

[chauhan7]

Significant values in the correlogram of Q statistics

[startz]

You have a large amount of data. That probably means strong evidence of a very small amount of autocorrelation.

[chauhan7]

I see so am I best to follow the DW stat? Also, what about the ARCH effect?

[startz]

I'm not sure that the Durbin-Watson is technically valid, but it's probably approximately right.

[chauhan7]

ok Thanks, so how do i go about removing the ARCH effect from the data?

[startz]

You might try allowing for higher-order GARCH effects. But I'm not sure why it's terribly important to get the variance equation exactly right. I guess it depends on what you're trying to do with it.

[chauhan7]

You might try allowing for higher-order GARCH effects. But I'm not sure why it's terribly important to get the variance equation exactly right. I guess it depends on what you're trying to do with it.

How do I allow for higher-order GARCH effects? It is needed to be correct to make sure that the interpretations I make from the data are as accurate as possible.

Thank you

[startz]

The estimate tab in the equation object lets you pick the order of the ARCH and GARCH terms.

[chauhan7]

Hm, thanks for that. So in my full sample I have used GARCH (2, 2) in addition to the ar(1) term to tackle both autocorrelation and the arch effect. What are the implications of this model in contrast to those of GARCH (1, 1). I cant seem to find a great deal of information on higher order models on the internet.