Exploring LLM Cryptocurrency Trading Through Fact-Subjectivity Aware Reasoning

TL;DR

The paper addresses why stronger LLMs do not always outperform weaker models in cryptocurrency trading by analyzing their reasoning processes. It introduces FS-ReasoningAgent, a multi-agent framework that splits reasoning into fact-based and subjective streams and integrates them through specialized agents and a trading core. Through experiments on BTC, ETH, and SOL across bull and bear markets, the approach yields higher profits and better risk-adjusted returns than CryptoTrade baselines and remains competitive with traditional strategies. A key finding is that subjective cues dominate in bull markets, while factual analysis drives bear-market performance, suggesting that harnessing advanced reasoning requires task-specific architectural design to unlock potential in financial AI planning.

Abstract

While many studies show that more advanced LLMs excel in tasks such as mathematics and coding, we observe that in cryptocurrency trading, stronger LLMs sometimes underperform compared to weaker ones. To investigate this counterintuitive phenomenon, we examine how LLMs reason when making trading decisions. Our findings reveal that (1) stronger LLMs show a preference for factual information over subjectivity; (2) separating the reasoning process into factual and subjective components leads to higher profits. Building on these insights, we propose a multi-agent framework, FS-ReasoningAgent, which enables LLMs to recognize and learn from both factual and subjective reasoning. Extensive experiments demonstrate that this fine-grained reasoning approach enhances LLM trading performance in cryptocurrency markets, yielding profit improvements of 7\% in BTC, 2\% in ETH, and 10\% in SOL. Additionally, an ablation study reveals that relying on subjective news generates higher returns in bull markets, while focusing on factual information yields better results in bear markets. Code is available at https://github.com/Persdre/FS-ReasoningAgent.

Figures (8)

• Figure 1: Comparison of Reasoning Processes - Trading Decisions Using News Data Alone; With/Without Fact and Subjectivity Agents on April 18, 2023 in the ETH Market, comparing GPT-3.5-turbo and o1-mini. The floating-point numbers represent buy/sell actions, where 0.7 indicates using 70% of available cash to buy ETH, and -0.3 indicates selling 30% of held ETH.
• Figure 2: Fact-Subjectivity Reasoning Agent Framework. This framework contains the following agents: Statistics Agent, Fact Agent, Subjectivity Agent, Fact Reasoning Agent, Subjectivity Agent, Trade Agent, and Reflection Agent. We provide an example of each agent's analysis displayed besides the corresponding agent.
• Figure : Performance comparison of single LLMs, and baseline trading strategies on ETH during both Bull and Bear market conditions.
• Figure : Performance of each strategy on BTC under both bull and bear market conditions. For each market condition and metric, the best result is highlighted in bold, the runner-up is indicated with an underline, and the best result among each families of LLM-based strategies is highlighted in green.
• Figure : Performance of each strategy on ETH under bull and bear market conditions.