Time series is not enough: Financial Transformer Reinforcement Learning for portfolio management

Existing portfolio management research is primarily focused on analyzing the relationships between price fluctuations of individual assets to optimize trading strategies. Most reinforcement learning research have several limitations in extracting inter-asset correlations, such as the inability to effectively simulate long-term market dynamics, interference in gradient propagation due to misaligned optimization objectives across sequentially connected modules, and difficulties in capturing complex spatiotemporal dependencies. These limitations lead to an overemphasis on local information during the feature extraction process, ultimately hindering the framework's ability to achieve global optimization and exacerbating the risk of overfitting. To address these limitations, we propose an innovative deep reinforcement learning framework, termed Financial Transformer Reinforcement Learning (FTRL). We evaluate the efficacy of FTRL against three learning variants, fourteen traditional strategies, and nine reinforcement learning methods across four Dow Jones and four NASDAQ datasets. Our results show that FTRL consistently outperforms all strategies in return, Sharpe ratio, and Sortino ratio, with at least 3.9%, 1.5%, and 5.3% improvements on Dow Jones datasets, respectively. On the NASDAQ datasets, FTRL achieves at least 40% higher average cumulative returns. We further validate the generalization of FTRL on the Greek stock market, where it delivers at least a 3.7% return advantage over other strategies, highlighting its applicability in emerging markets.