ARCH & GARCH volatility

[manalyst]

Hi all,

I am trying to model exchange rate volatility in the MYR/USD pair.

So i created a logged returns series from the nominal exchange rates. Then I found that fitting an ar(5) model eliminates the autocorrelation in the residuals in the mean equation, while there still remains autocorrelation in the squared residuals (clustering). Next would be fitting the variance equation. However, when I fit an ARCH/GARCH/any variance equation, the previously eliminated autocorrelation in the un-squared residuals come back..

Why does this happen and how can i fix it?

Thanks,

[trubador]

Time varying volatility may induce serial correlation in the mean equation. So you could have observed otherwise: autocorrelation might have dissappeared after building an explicit variance model. You can continue to add AR(p) terms into the mean part until you remove the serial correlation, while controlling the changing variance (i.e. GARCH part).