The paper evaluates the contribution of conditional second moments, from high-frequency data, to optimal portfolio allocation. Using the DCC model as a benchmark, we put forth two novel approaches: a model for the inverse conditional correlation matrix (DCIC) and the direct modeling of the conditional portfolio weights (DCW). We assess their out-of-sample ability by comparing the corresponding minimum-variance portfolios built on the components of the Dow Jones 30 Index. Evaluating performance in terms of portfolio variance, certainty equivalent, turnover and break-even transaction costs, we find that exploiting conditional correlations gives marked improvements upon volatility timing and na ̈ıve strategies: DCC and the computationally convenient DCIC perform in a similar way; DCW, the simplest and fastest to implement, exhibits either equal or superior performances with respect to the measures considered.