Re: [Gretl-users] OPG inversion problem

On Sun, 20 Nov 2011, Dan Běsoň wrote: > > Dear everybody, > > I am trying to estimate an ARMA process using MLE and then to extend > it by jointly estimating ARMA-GARCH. The following code [...] > mle ll=-0.5*(log(h) + (e^2)/h) > series e = in - mu > params c phi1 phi2 phi3 phi4 phi5 phi6 phi7 theta1 theta2 theta3 > theta4 theta5 theta6 > end mle --hessian > > produces an "failed to invert OPG matrix GG" error regardless of > whether I use --hessian. Your loglikelihood is a function of h and e; e is a function of mu. None of these are functions of the parameters you want to optimise on. Consequently, the score matrix contains all zeros and its cross product is a square matrix of zeros, which is notoriously hard to invert.