This paper introduces CryptoTrade, a novel LLM-based trading agent specifically designed for cryptocurrency markets. Unlike previous LLM applications primarily focused on stock trading, this work addresses the unique characteristics of cryptocurrency markets by integrating both on-chain data (leveraging blockchain transparency and immutability) and off-chain signals (such as news and social media influence). The agent incorporates a reflective mechanism that continuously refines its trading decisions by analyzing the outcomes of previous trades, enabling adaptive learning and improved performance over time.

The research makes two significant contributions: expanding LLM applications to cryptocurrency trading and establishing a benchmark for cryptocurrency trading strategies. Through comprehensive experiments across various cryptocurrencies and market conditions, CryptoTrade demonstrates superior performance in maximizing returns compared to both traditional trading approaches and time-series baselines. The availability of code and data facilitates reproducibility and further research in this emerging field of LLM-driven cryptocurrency trading.