A machine learning approach to the optimal estimation of portfolio weights

The purpose of this thesis is to review and expand the main result in the paper by Daniel Kinn, "Reducing Estimation Risk in Mean-Variance Portfolios with Machine Learning (preprint arXiv:1804.01764 (2018)). This result states that maximizing a quadratic utility function for portfolio optimization is equivalent to minimizing the generalization error of a machine learning algorithm. This assertion is then applied to Linear Models and Neural Networks. Finally, the exposed theoretical approaches are coded and tested with real data, generating several portfolio optimization algorithms whose predictions are then commented upon and compared with the classical methods. Code and new insights for handling current hazards are made available for further exploration.Outgoin