The RATS Software Forum

Hello Tom,

another quick question:

How does RATS handle the starting values for the coefficients if I use the GARCH instruction for a multivariate model (with mean equations specified using "group"), but without the "initial=" option. Does it somehow generate its own starting / guess values or does it just use a %beta vector with all zeros, which it then, for example, improves upon with the simplex and then the BFGS?

I'm estimating a model with a multivariate t, and I thought that starting the estimation with a guess value for the scale parameter that is equal to zero should probably lead to some problems right from the beginning of the estimation. So my guess is that RATS somehow comes up with a different number as a starting value. The question is just which one and what about the other coefficients?

For a mean model other than just the default constant, the guess values are 0 which usually works fine. (The estimation is more sensitive to guesses for the GARCH parameters.) You definitely don't want to start with a guess of 0 for the shape, since the likelihood isn't even defined there. For the t, the guess is 20, so it's guessed to be slightly fatter than Normal.