Solving dynamic Portfolio optimization through Reinforcement Learning and Recurrent Neural Networks

Portfolio optimization is a common task both for institutional and retail investors. Since Artificial Intelligence applications are spreading in the finance industry, this thesis addresses how to harness AI decision-making models to provide advice for investors. Quantitative models for portfolio management mostly rely on the Modern Portfolio Theory with proven profits but also constraints. Inspired by recent literature on Reinforcement Learning (RL) agents as Robo advisors, the thesis developed a tailor-made RL model for dynamic portfolio optimization enhanced by a recurrent neural network (RNN). Modern Portfolio Theory, even if succeed in maximizing the profit at a given level of risk, has stationary assumptions and fails especially in high volatility markets. On the contrary, Reinforcement Learning offers a deep understanding of price behaviour adapting its policy over specific features of the assets. The thesis finally tests the dominance of RL along with RNN over a RL with a more basic design and the Maximum Sharpe Portfolio picked from the efficient frontier.