Robust Portfolio Optimization with Correlation Penalties

Sammanfattning: Robust portfolio optimization models attempt to address the standard optimization method's high sensitivity to noise in the parameter estimates, by taking an investor's uncertainty about the estimates into account when finding an optimal portfolio. In this thesis, we study robust variations of an extension of the mean-variance problem, where an additional term penalizing the portfolio's correlation with an exogenous return sequence is included in the objective. Using a normalized risk factor model of the asset returns, estimations are done using EMA filtering as well as exponentially weighted linear regression. We show that portfolio performance can significantly improve with respect to a range of metrics, such as Sharpe ratio, expected shortfall and skewness, when using appropriate robust models and hyperparameters. We further show that extending the optimization problem with a correlation penalty can notably reduce portfolio correlation with an arbitrary return sequence, with only a small impact on other performance metrics.