ИСТИНА

We explore the effectiveness of various machine learning algorithms, especially deep reinforcement learning, for solving the portfolio optimization problem. The investigated algorithms can be divided into the following groups: Follow-the-Winner (using trend follow principle), Follow-the-Loser (using mean reversion principle) Pattern-Matching (for example, using correlations between current prices and historical prices), Meta-Learning Algorithms (combination of several algorithms) and, finally, strategies based on reinforcement learning. We propose a novel meta-learning algorithm based on deep reinforcement learning, which we call MetaRL. It utilizes algorithms from different groups and optimizes their combination. Reinforcement learning (RL) technique is well suited to solve portfolio selection tasks because it can learn hidden dependencies in financial data and trading patterns by trial and error, directly predicts trading actions, and allows to consider delayed rewards. Moreover, RL naturally takes into account transaction costs and can model real trading when the strategy can impact the market. As agents, we use deep neural networks that decide which of the algorithms will be more profitable at a particular moment. We provide extensive experiments using financial data and show that our new MetaRL algorithm substantially outperforms other portfolio optimization methods.