Risk-Averse Reinforcement Learning for Portfolio Optimization

We investigated investment-portfolio optimization using Reinforcement Learning (RL) with risk assessment. Due to market friction, the reaction of other market participants and uncertainties, it is challenging to trade and optimize investment portfolios dynamically. The financial market is sophisticated and complex to model. Moreover, regulatory requirements and internal risk policy require investors to make risk-averse decisions to prevent catastrophic results. One way to solve the problem is to set a high enough penalty to reward for making a risky decision beyond investor’s risk appetite. We utilized Bayesian Neural network (BNN) which can yield not only the value function, but also it provides uncertainty on less experienced or inherent stochastic states. We have successfully decreased the risk/uncertainties in the agent training. The risk-averse RS algorithm showed 18 percent less riskiness during the training process. We conclude that implementing a risk-averse RL technique could be used portfolio optimization by risk-averse investors