Mitigating Risk in Machine Learning-Based Portfolio Management: A Deep Reinforcement Learning Approach

Many researchers when using machine learning approaches for investing focus on predicting returns, but overlook risk management or portfolio allocation. That may lead to risk management problems or even uncontrolled strategy risk. In order to address such vulnerability, we propose a framework for training an allocation strategy that models the investment process as a Markov decision process. We use a risk decomposition technique and a deep reinforcement learning agent model to compute variances and update portfolio variables efficiently. We tested our framework with a LSTM-based agent for predicting stock movements and found that it was able to offset systematic risk and achieve reasonable returns. The proposed low-dimensional framework may contribute to more effective portfolio management based on reinforcement learning.