Covariance Matrix Forecasting: Average Oracle Method
Roman R.
Continuing this series on covariance matrix forecasting (c.f. here and there for the previous posts), I will now describe a relatively recent1 data-driven, model-free, way to [forecast] covariance [and correlation] matrices of time-varying systems2 rooted in random matrix theory. This method - introduced in Bongiorno et al.2 and called Average Oracle - consists in replacing the eigenvalues of a (noisy) estimate of a time-varying covariance matrix by time-independant eigenvalues that encode the a