fmincon hessian problem
in [Matlab]
|
From: Manthos Vogiatzoglou on 28 Feb 2010 18:55 I do not know if this would help but when I had a similar problem Peter Perkins from mathworks suggested that the numerical Hessian provided by the fmincon is not that good approximation of the real hessian anlike fminunc which produces a better approximation. In the case of the normal copula you are interested in, you could use a new variable, say q (any real number) and r = atan(2p/pi) for the correlation coefficient (in -1 ,1). In that way you can reparametrize your problem and use fminunc instead. In Kevin Sheppard's home page you can find some codes for calculating numerical gradient and hessian. In a typical copula context the margins are fitted to a GARCH type model first, prior to the copula fitting. If that is your case you should consider sandwich (or bootstrap) estimators for the sandard errors.
|
Pages: 1 Prev: Function call from different directory Next: MATLAB SOURCE CODE FOR JPEG2000 |