This paper addresses the debate about the usefulness of high-frequency (HF) data in large-scale portfolio allocation. We construct global minimum variance portfolios based on the constituents of the S&P 500. HF-based covariance matrix predictions are obtained by applying a blocked realized kernel estimator, different smoothing windows, various regularization methods and two forecasting models. We show that HF-based predictions yield a significantly lower portfolio volatility than methods employing daily returns. Particularly during the 2008 financial crisis, these performance gains hold over longer horizons than previous studies have shown, translating into substantial utility gains for an investor with pronounced risk aversion.
