An enhanced deep reinforcement learning ensemble empowered by large language model

Abstract

The domain of financial trading operates at the intersection of diverse informational inputs, including asset pricing and technical indicators. These elements inform critical financial activities such as quantitative trading across varied asset classes. Despite the growing use of advanced Artificial Intelligence (AI) in finance, applying these technologies to trading tasks often runs into challenges due to insufficient analysis of diverse formats of data and the ability to adapt to diverse trading scenarios. This study introduces a framework that synergizes an ensemble of deep reinforcement learning algorithms with the explanatory power of a Large Language Model. The framework termed ”Large Language Model-Enhanced Deep Reinforcement Learning Ensemble” (LLM-DRE), stands at this crossroad, ingeniously combining the robust decision-making capabilities of DRL with the nuanced contextual understanding of LLMs. At its core, LLM-DRE effectively processes and synthesizes disparate forms of market data to fine-tune trading strategies, enabling an enriched assessment of financial environments. Performance-wise, LLM-DRE not only holds its ground but excels against traditional baselines, exhibiting substantial enhancements in profitability and risk-adjusted returns. It systematically outperforms the benchmarks across various assets, demonstrating particularly impressive results on volatile markets, where its adaptive strategies foresee and leverage rapid price movements to secure profitable gains.