Initialization in markov switching model

[hakeru7]

Hello,

I understand that the initial coefficients in a markov switching estimation is created by doing a simple ols and adding random normals around the coefficients using their respective s.d.

I'm wondering how the initial regime specific error variances are set and how the initial probability parameters set as well.

Thanks!

[EViews Glenn]

Completely arbitrary rule of taking the log of the sigma obtained from OLS, and adding a randomized deviation based on the original log sigma, the specified search scale, and a random normal.
Probability regressors initialized to zero.

[hakeru7]

Thanks.

Also, I don't think I found a description of what algorithm the estimation uses in numerically maximizing the likelihood. Could you hint?

[EViews Glenn]

BFGS in EViews 8.

BFGS default in EViews 9. Also options for OPG/BHHH and Newton.