<italic>HADAPS</italic>: Hierarchical Adaptive Multi-Asset Portfolio Selection

Abstract

Multi-asset portfolio selection is an asset allocation strategy involving a variety of assets. Adaptive investment strategies which consider the dynamic market characteristics of individual assets and asset classes are vital for maximizing returns and minimizing risks. We introduce HADAPS, a novel computational method for multi-asset portfolio selection which utilizes the Soft-Actor-Critic (SAC) framework enhanced with Hierarchical Policy Network. Contrary to previous approaches that have relied on heuristics for constructing asset allocations, HADAPSdirectly outputs a continuous vector of action values depending on current market conditions. In addition, HADAPS performs multi-asset portfolio selection involving multiple asset classes. Experimental results show that HADAPS outperforms baseline approaches in not only cumulative returns but also risk-adjusted metrics. These results are based on market price data from sectors with various behavioral characteristics. Furthermore, qualitative analysis shows HADAPS ’ ability to adaptively shift portfolio selection strategies in dynamic market conditions where asset classes and different assets are uncorrelated to each other.

Keywords

• Portfolios
• investment
• stock markets
• cryptocurrency
• reinforcement learning