Deep reinforcement learning for portfolio management of markets with a dynamic number of assets

Abstract

This work proposes a novel portfolio management method using deep reinforcement learning on markets with a dynamic number of assets. This problem is especially important in cryptocurrency markets, which already support the trading of hundreds of assets with new ones being added every month. A novel neural network architecture is proposed, which is trained using deep reinforcement learning. Our architecture considers all assets in the market, and automatically adapts when new ones are suddenly introduced, making our method more general and sample-efficient than previous methods. Further, transaction cost minimization is considered when formulating the problem. For this purpose, a novel algorithm to compute optimal transactions given a desired portfolio is integrated into the architecture. The proposed method was tested on a dataset of one of the largest cryptocurrency markets in the world, outperforming state-of-the-art methods, achieving average daily returns of over 24%.