TLDR: A new research paper introduces an adaptive multi-agent Bitcoin trading system that utilizes Large Language Models (LLMs) for alpha generation and portfolio management. The system employs specialized agents for technical analysis, sentiment evaluation, and decision-making. Its key innovation is a dual verbal feedback mechanism, where a ‘Reflect’ agent provides daily and weekly natural-language critiques of trading decisions. These critiques are integrated into future prompts, enabling the system to refine its strategies without traditional parameter updates or fine-tuning. Back-testing results show consistent outperformance against buy-and-hold strategies and significant improvements in profitability and robustness, demonstrating verbal feedback as an effective and low-cost method for optimizing LLMs in finance.
Trading Bitcoin, with its notorious volatility and susceptibility to rapid shifts in market sentiment and regulatory news, has always been a challenging endeavor. Traditional models, often reliant on static historical data, struggle to keep pace with the dynamic nature of the cryptocurrency market. A new research paper, An Adaptive Multi-Agent Bitcoin Trading System, introduces an innovative approach to tackle this problem by leveraging Large Language Models (LLMs) within a multi-agent framework.
Authored by Aadi Singhi from UCL Computer Science, this paper outlines a system designed for alpha generation and portfolio management specifically for Bitcoin. The core idea is to structure LLMs into specialized agents, each focusing on a distinct aspect of market analysis and decision-making.
The Specialized Agents
The system comprises several key agents:
-
Quantitative Agent (Quants): This agent performs technical analysis, processing historical and real-time market data such as price trends, moving averages, MACD, RSI, and on-chain metrics like transaction volumes. Its role is to classify the next day’s market state (bullish, bearish, or neutral) and suggest a trading strategy based purely on technical insights.
-
Sentiment Agent (Signals): Focused on market psychology, this agent assesses investor sentiment. It gathers data from financial news outlets, the Fear and Greed Index, and social media sentiment scores to understand public mood and external influences on Bitcoin’s price.
-
Decision Agent: Acting as the central intelligence, the Decision Agent consolidates the predictions and reasoning from both the Quants and Signals agents. It then formulates a final market prediction and a portfolio allocation strategy, aiming to outperform a static 50% Bitcoin-50% cash baseline.
The Power of Verbal Feedback
What truly sets this system apart is its novel dual verbal feedback mechanism. Unlike conventional methods that require complex parameter updates or fine-tuning, this system learns and adapts through natural-language critiques. This mechanism involves two components:
-
Reflect Agent (Daily Feedback): This agent provides daily performance reports and suggestions. It evaluates the qualitative reasoning of the Quants, Signals, and Decision agents against actual market movements and performance metrics. Crucially, it critiques whether decisions were justified by the reasoning provided, rather than just the outcome, helping agents identify flaws in their analytical process.
-
Long-Term Reflect Agent (Weekly Feedback): To address the short-term memory limitations of daily feedback, this agent provides weekly evaluations. It condenses a full trading week’s performance into a summary, offering standardized guidance like “prioritize high-confidence inputs” or “reconstruct your indicator selection.” This weekly context helps agents maintain a broader view and refine their reasoning over time.
This verbal feedback is then injected into future prompts, allowing the agents to adjust their indicator priorities, sentiment weights, and allocation logic without any direct parameter changes. This makes the system highly flexible, transparent, and cost-efficient for dynamic financial environments.
Also Read:
- CGA-Agent: A Hybrid Framework for Adaptive Crypto Trading Strategy Optimization
- Unmasking the “Profit Mirage” in AI Financial Trading
Impressive Results
Back-testing on Bitcoin price data from July 2024 to April 2025 demonstrated consistent outperformance across various market conditions. The Quantitative agent delivered over 30% higher returns in bullish phases and a 15% overall gain compared to a buy-and-hold strategy. The sentiment-driven agent transformed sideways markets from a small loss into a gain of over 100%. The introduction of weekly feedback further boosted total performance by 31% and reduced losses during bearish periods by 10%.
These results highlight that verbal feedback is a scalable and low-cost method for enhancing LLMs for financial goals, allowing them to adapt dynamically and replicate the benefits of human oversight without additional training costs. The system’s focus on profitability over mere predictive accuracy aligns its objectives more closely with real-world investment outcomes, marking a significant step forward in AI-driven cryptocurrency trading.
