Modern Approaches to Dynamic Portfolio Optimization

Abstract

Although appealing from a theoretical point of view, empirical assessments of dynamic portfolio optimizations in a mean-variance framework often fail to reach the high expectations set forth by analytical evaluations. A major reason for this shortfall is the imprecise estimation of asset moments and in particular, the expected return. This work levers recent advancements in the field of machine learning and employs three types of artificial neural networks in an attempt to improve the accuracy of the asset return estimation and the expected associated portfolio performances. After an introduction of the dynamic portfolio optimization framework and the artificial neural networks, their suitability for the considered application is analyzed in a two asset universe of a market and a risk-free asset. A comparison of the corresponding risk-return characteristics and those achieved using a more traditional exponentially weighted moving average estimator is subsequently drawn. While outperformance of the artificial neural networks is found for daily and monthly estimated returns, significance can only be established in the latter case, especially in light of trading costs. Multiple robustness checks are performed before an outlook for subsequent research opportunities is given.

Keywords: Portfolio optimization; machine learning; multilayer perceptron; convolutional neural network; long short-term memory.