Cleaning the covariance matrix of strongly nonstationary systems with time-independent eigenvalues

We propose a data-driven way to clean covariance matrices in strongly nonstationary systems. Our method rests on long-term averaging of optimal eigenvalues obtained from temporally contiguous covariance matrices, which encodes the average influence of the future on present eigenvalues. This zero-th order approximation outperforms optimal methods designed for stationary systems.