Estimation of Covariance Matrices for Portfolio Optimization using Gaussian Processes.

Estimating covariances between financial assets plays an important role in risk management and optimal portfolio allocation. In practice, when the sample size is small compared to the number of variables, i.e. when considering a wide universe of assets over just a few years, this poses considerable challenges and the empirical estimate is known to be very unstable. Here, we propose a novel covariance estimator based on the Gaussian Process Latent Variable Model (GP-LVM). Our estimator can be considered as a non-linear extension of standard factor models with readily interpretable parameters reminiscent of market betas. Furthermore, our Bayesian treatment naturally shrinks the sample covariance matrix towards a more structured matrix given by the prior and thereby systematically reduces estimation errors.