New ways to optimize customers portfolios with Reinforcement Learning
In financial markets future status are driven by actions of many decision makers. Reinforcement Learning (RL) is a Machine Learning subfield in which an agent interacts with the environment in the setting of a sequential decision-making process. The agent chooses actions, among many possible, at each step of such a process in order to maximize a total cumulative reward. Portfolio optimization is an ideal candidate for an RL-based approach: when constructing a portfolio, the aim is to maximize returns while minimizing the risk with respect to the amount of money we allocate on each asset in the portfolio in a sequential decision-making process.
Insight & Action
The RL approach to the solving portfolio optimization problem would proceed accordingly to the following steps:
- an AI agent receives information about the current state of the of the market
- performs an action on the market by allocating weights in the portfolio
- these actions, together with other external actions, brings the market environment in a new state
- the AI agent receives a reward based on the portfolio value after the weights application
- this cycle is repeated with the goal of learning a control policy of the system that maximizes the total expected lifetime reward
Knowledge Lab chooses a RL approach for portfolio optimization instead of the standard portfolio management approaches, as with RL we are able to generalize strategies across assets and markets regardless of the training universe. The RL approach has been proven to outperform the state-of-art portfolio management approaches (see, e.g., Filos A., Reinforcement Learning for Portfolio Management, ArXiv, abs/1909.09571 (2019)).