Forecasting Realized Correlations : A MIDAS Approach

Mixed data sampling (MIDAS) regression has received much attention in relation to modeling financial time series due to its flexibility. Previous work has mainly focused on forecasting of realized volatilities and has rarely been used to predict realized correlations. This paper considers a MIDAS approach to forecast realized correlation matrices. A MIDAS model is estimated via nonlinear least squares (NLS) using an analytical gradient-based optimization. Based on the model confidence set (MCS) procedure we discover that the introduced approach is superior compared to the established heterogeneous autoregressive (HAR) model in terms of out-of-sample forecasting accuracy. This preeminence is due to the flexible data-driven origin of the MIDAS model. The latter results in higher economic value with regard to portfolio management applications. The improvement is considerable for longer forecasting horizons both in calm times and during the periods of market turbulence