The classical mean-variance portfolio model was originally proposed by Markowitz (1952). It has now undergone 65 years of development. In the mean-variance portfolio model, the mean and the covariance matrix of asset returns are often unknown and need to be estimated. However, the sampling errors have adverse effects on portfolio performance, leading to sub-optimal and unstable portfolio weights. Various strategies have been proposed to reduce the sampling errors. In this talk, both the traditional methods and some modern high-dimensional statistical approaches are widely reviewed. Moreover, a new approach based on the shrinkage of the sample eigenvalues is proposed, aimed at reducing the over-dispersion issue of the sample eigenvalues. The empirical studies show that the proposed approach can often achieve a lower out-of-sample variance and higher Sharpe ratio than the existing portfolio strategies in most real data sets.