Volatility Modelling- r(430) error & Negative Coefficients

[stefan90]

Dear Statisticians,

I am having some issues modeling the effect of news events on market indices. Namely, I am using garch (1,1) and have 8 dummy variables each corresponding to different dates.

Ideally, I would like to have something that looks like the following: arch lnfdbra, arch(1/1) garch(1/1) het(Lehman Usbail Icebail Irbail Portbail Spabail Grkbail Cypbail)

The issue is that I get an error message "could not calculate numerical derivatives missing values encountered". r(430)

My procedure so far has been as follows:
First I checked the stationarity using the ADF test. As a result, I took the log first difference of the data and used the akaike information criteria to determine which model best captured the volatility. Garch (1,1) was best and next I tested the residuals for white noise, which indicated that the model was fine. Adding the dummy variable for Lehman works perfectly. However, as soon as I add another variable, I get the error message.

For example:
This works: arch lnfdbra, arch(1/1) garch(1/1) distribution(t) het(Lehman)
This doesn't: arch lnfdbra, arch(1/1) garch(1/1) distribution(t) het(Lehman Usbail Portbail )

When I run the regression with the dummies one by one, I either get good results, negative coefficients, or the error message. How is it possible for the coefficients to be negative (sometimes the values are -1000) and why does the error message appear? Does anybody have any suggestions/solutions to this problem? I would really appreciate any help!

[Jorge Eduardo Perez Perez]

Wouldn't negative coefficients mean that the event in question reduces the conditional variance? Statistically these are not necessarily wrong unless the negative coefficient is large enough to make the variance negative.

My limited experience with arch tells me that the estimation is very difficult and sensitive to optimization options. Have you tried changing those? See maximization_options under help arch

[Jorge Eduardo Perez Perez]

Oh and by the way, Statalist courtesy is to post under your real name. Please read the FAQ. Also please provide tags for your question. Suggested tags are: Time Series, Garch.