Estimation noise is a well-known issue in empirical portfolio modelling. However, existing models suffer from large forecasting errors which dominate the theoretical gain. In this paper, we propose a direct weight estimator (DWE), which accounts for forecasting risk and avoids the over-parametrization problem by forecasting a one-dimensional portfolio measure directly. We define a forecasting error based on realized measures and optimize for a weight vector which results in a more precise forecast and at the same time is not far from the optimal portfolio solution. The DWE is shown to outperform commonly used approaches in both simulation and empirical studies.