• contextual bandit
• portfolio management
• reinforcement learning

Market exposed assets like stocks yield higher return than cash but have higher risk, while cash-equivalent assets yield little to no risk, but their value can only decrease in time. Portfolio management consists in using predictions of the market’s price change to decide the percentage of market exposed assets held in a portfolio. We identify two problems that can be improved using reinforcement learning. Reinforcement learning is a subclass of machine learning where an agent takes actions to interact with an environment. Actions are chosen following a policy that maps state of the environment to probability distributions on the actions' space. After each action the agent receives a reward signal and the environment transitions to a new state. The tuple formed by the old state, action, reward and new state is used to improve the policy to maximize the cumulative sum of rewards. The first problem considered in this work comes from using deep learning methods to predict price changes, that produce different predictions due to model initialization with no clear way to decide which is the best. We use Multi Armed Bandits, reinforcement learning algorithms for one-state environments, to predict the best performing initialization. The second problem is the use of a fixed policy to decide the market exposure, which is unrelated to the particular portfolio to manage. We propose the use of a reinforcement learning agent to derive portfolio-specific policies.