When: Dec 9th, 2021 – 11:00 – 11:45 AM
Where: Google meet link
One of the most studied problems in the financial framework is how to find a combination of N assets, called Portfolio, which optimizes an investment objective. This problem is called “Portfolio Optimization Problem”. An optimal solution for this problem is called “Optimal Portfolio”, which is usually intended as the least risky one. In this Master Thesis work, I try to solve the Portfolio Optimization Problem making use of Reinforcement Learning (RL). The goal is to build a dynamic portfolio, which selects independently the fractions of capital to invest on the N assets for minimizing the risk of loss. In order to do so, I implement an Actor-Critic (AC) algorithm in Pytorch for rebalancing the portfolio in an optimal way, observing the returns of the assets at the previous time step. Making use of Deep Neural Networks (DNNs), the “AC Portfolio” is able to learn a stochastic optimal policy: the agent samples actions from a probability density function (PDF) and learns its parameters through training. In this way, the “AC Portfolio” is able to find the optimal financial strategy observing the market data, without referring to any underlying model for describing the behavior of the returns. Due to the stochasticity of the actions and the returns of the assets (which cannot be controlled directly by the agent), the AC Portfolio has a high variance in the results, which can be reduced using different algorithms, like the Soft-Actor Critic (SAC) algorithm.