This paper discusses the asymptotic efficiency of estimators for optimal portfolios when the returns are vector-valued Gaussian stationary processes. Then it is shown that the usual portfolio estimators are not asymptotically efficient if the returns are dependent. Numerical studies for the difference between the asymptotic variance of the portfolio estimators and the Cramer-Rao bound are given. The results clearly illuminate the inefficiency of the usual estimators for vector-valued ARMA(1,2) processes. >From this point of view we construct portfolio estimators which are asymptotically efficient.