Why do mean-variance (MV) models perform so poorly? In searching for an answer to this question, we estimate expected returns by sampling from a multivariate probability model that explicitly incorporates distributional asymmetries. Specifically, our empirical analysis shows that an application of copulas using marginal models that incorporate dynamic features such as autoregression, volatility clustering, and skewness to reduce estimation error in comparison to historical sampling windows. Using these copula-based models, we find that several MV-based rules exhibit statistically significant and superior performance improvements even after accounting for transaction costs. However, we find that outperforming the naïve equally-weighted (1/N) strategy after accounting for transactions costs still remains an elusive task.