TradingAgents: Multi-Agents LLM Financial Trading Framework

Autonomous agents leveraging Large Language Models (LLMs) present a transformative approach to decision-making by replicating human processes and workflows across various applications. These systems enhance the problem-solving capabilities of language agents by equipping them with tools and enabling collaboration with other agents, effectively breaking down complex problems into manageable components. One prominent application of these autonomous frameworks is in the financial market—a highly complex system influenced by numerous factors, including company fundamentals, market sentiment, technical indicators, and macroeconomic events.

Traditional algorithmic trading systems often rely on quantitative models that struggle to fully capture the complex interplay of diverse factors. In contrast, LLMs excel at processing and understanding natural language data, making them particularly effective for tasks that require textual comprehension, such as analyzing news articles, financial reports, and social media sentiment. Additionally, deep learning-based trading systems often suffer from low explainability, as they rely on hidden features that drive decision-making but are difficult to interpret. Recent advancements in multi-agent LLM frameworks for finance have shown significant promise in addressing these challenges. These frameworks create explainable AI systems, where decisions are supported by evidence and transparent reasoning, demonstrating their potential in financial applications.

Despite their potential, most current applications of language agents in the financial and trading sectors face two significant limitations:

Lack of Realistic Organizational Modeling: Many frameworks fail to capture the complex interactions between agents that mimic the structure of real-world trading firms. Instead, they focus narrowly on specific task performance, often disconnected from the organizational workflows and established human operating procedures proven effective in trading. This limits their ability to fully replicate and benefit from real-world trading practices. Inefficient Communication Interfaces: Most existing systems use natural language as the primary communication medium, typically relying on message histories or an unstructured pool of information for decision-making. This approach often results in a "telephone effect", where details are lost, and states become corrupted as conversations lengthen. Agents struggle to maintain context and track extended histories while filtering out irrelevant information from previous decision steps, diminishing their effectiveness in handling complex, dynamic tasks. Additionally, the unstructured pool-of-information approach lacks clear instructions, forcing logical communication and information exchange between agents to depend solely on retrieval, which disrupts the relational integrity of the data.

In this work, we address these key limitations of existing models by introducing a system that overcomes these challenges. First, our framework bridges the gap by simulating the multi-agent decision-making processes typical of professional trading teams. It incorporates specialized agents tailored to distinct aspects of trading, inspired by the organizational structure of real-world trading firms. These agents include fundamental analysts, sentiment/news analysts, technical analysts, and traders with diverse risk profiles. Bullish and bearish debaters evaluate market conditions to provide balanced recommendations, while a risk management team ensures that exposures remain within acceptable limits. Second, to enhance communication, our framework combines structured outputs for control, clarity, and reasoning with natural language dialogue to facilitate effective debate and collaboration among agents. This hybrid approach ensures both precision and flexibility in decision-making.

We validate our framework through experiments on historical financial data, comparing its performance against multiple baselines. Comprehensive evaluation metrics, including cumulative return, Sharpe ratio, and maximum drawdown, are employed to assess its overall effectiveness.

Large Language Models (LLMs) are applied in finance by fine-tuning on financial data or training on financial corpora. This improves the model’s understanding of financial terminology and data, enabling a specialized assistant for analytical support, insights, and information retrieval, rather than trade execution.

Fine-Tuned LLMs for Finance

Fine-tuning enhances domain-specific performance. Examples include PIXIU (FinMA), which fine-tuned LLaMA on 136K finance-related instructions; FinGPT, which used LoRA to fine-tune models like LLaMA and ChatGLM with about 50K finance-specific samples; and Instruct-FinGPT, fine-tuned on 10K instruction samples from financial sentiment analysis datasets. These models outperform their base versions and other open-source LLMs like BLOOM and OPT in finance classification tasks, even surpassing BloombergGPT in several evaluations. However, in generative tasks, they perform similarly or slightly worse than powerful general-purpose models like GPT-4, indicating a need for more high-quality, domain-specific datasets.

Finance LLMs Trained from Scratch

Training LLMs from scratch on finance-specific corpora aims for better domain adaptation. Models like BloombergGPT, XuanYuan 2.0, and Fin-T5 combine public datasets with finance-specific data during pretraining. BloombergGPT, for instance, was trained on both general and financial text, with proprietary Bloomberg data enhancing its performance on finance benchmarks. These models outperform general-purpose counterparts like BLOOM-176B and T5 in tasks such as market sentiment classification and summarization. While they may not match larger closed-source models like GPT-3 or PaLM, they offer competitive performance among similar-sized open-source models without compromising general language understanding.

In summary, finance-specific LLMs developed through fine-tuning or training from scratch show significant improvements in domain-specific tasks, underscoring the importance of domain adaptation and the potential for further enhancements with high-quality finance-specific datasets.

LLMs act as trader agents making direct trading decisions by analyzing external data like news, financial reports, and stock prices. Proposed architectures include news-driven, reasoning-driven, and reinforcement learning (RL)-driven agents.

News-Driven Agents

News-driven architectures integrate stock news and macroeconomic updates into LLM prompts to predict stock price movements. Studies evaluating both closed-source models (e.g., GPT-3.5, GPT-4) and open-source LLMs (e.g., Qwen, Baichuan) in financial sentiment analysis have shown the effectiveness of simple long-short strategies based on sentiment scores. Further research on fine-tuned LLMs like FinGPT and OPT demonstrates improved performance through domain-specific alignment. Advanced methods involve summarizing news data and reasoning about their relationship with stock prices.

Reasoning-Driven Agents

Reasoning-driven agents enhance trading decisions through mechanisms like reflection and debate. Reflection-driven agents, such as FinMem and FinAgent, use layered memorization and multimodal data to summarize inputs into memories, inform decisions, and incorporate technical indicators, achieving superior backtest performance while mitigating hallucinations. Debate-driven agents, like those in heterogeneous frameworks and TradingGPT, enhance reasoning and factual validity by employing LLM debates among agents with different roles, improving sentiment classification and increasing robustness in trading decisions.

Reinforcement Learning-Driven Agents

Reinforcement learning methods align LLM outputs with expected behaviors, using backtesting as rewards. SEP employs RL with memorization and reflection to refine LLM predictions based on market history. Classical RL methods are also used in trading frameworks that integrate LLM-generated embeddings with stock features, trained via algorithms like Proximal Policy Optimization (PPO).

LLMs are also used to generate alpha factors instead of making direct trading decisions. QuantAgent demonstrates this by leveraging LLMs to produce alpha factors through an inner-loop and outer-loop architecture. In the inner loop, a writer agent generates a script from a trader's idea, while a judge agent provides feedback. In the outer loop, the code is tested in the real market, and trading results enhance the judge agent. This approach enables progressive approximation of optimal behavior.

Subsequent research, such as AlphaGPT, proposes a human-in-the-loop framework for alpha mining with a similar architecture. Both studies showcase the effectiveness of LLM-powered alpha mining systems, highlighting their potential in automating and accelerating the development of trading strategies by generating and refining alpha factors.

Assigning LLM agents clear, well-defined roles with specific goals enables the breakdown of complex objectives into smaller, manageable subtasks. Financial trading is a prime example of such complexity, demanding the integration of diverse signals, inputs, and specialized expertise. In the real world, this approach to managing complexity is demonstrated by trading firms that rely on expert teams to collaborate and make high-stakes decisions, underscoring the multifaceted nature of the task.

In a typical trading firm, vast amounts of data are collected, including financial metrics, price movements, trading volumes, historical performance, economic indicators, and news sentiment. This data is then analyzed by quantitative experts (quants), including mathematicians, data scientists, and engineers, using advanced tools and algorithms to identify trends and predict market movements.

Inspired by this organizational structure, TradingAgents defines seven distinct agent roles within a simulated trading firm: Fundamentals Analyst, Sentiment Analyst, News Analyst, Technical Analyst, Researcher, Trader, and Risk Manager. Each agent is assigned a specific name, role, goal, and set of constraints, alongside predefined context, skills, and tools tailored to their function. For example, a Sentiment Analyst is equipped with tools like web search engines, Reddit search APIs, X/Twitter search tools, and sentiment score calculation algorithms, while a Technical Analyst can execute code, calculate technical indicators, and analyze trading patterns. More specifically, TradingAgents assumes the following teams.

Analyst Team

Researcher Team

Trader Agents

Risk Management Team

Communication Protocol

Most existing LLM-based agent frameworks use natural language as the primary communication interface, typically through structured message histories or collections of agent-generated messages. However, relying solely on natural language often proves insufficient for solving complex, long-term tasks that require extensive planning horizons. In such cases, pure natural language communication can resemble a game of telephone—over multiple iterations, initial information may be forgotten or distorted due to context length limitations and an overload of text that obscures critical earlier details. To address this limitation, we draw inspiration from frameworks like MetaGPT, which adopt a structured approach to communication. Our model introduces a structured communication protocol to govern agent interactions. By clearly defining each agent's state, we ensure that each role only extracts or queries the necessary information, processes it, and returns a completed report. This streamlined approach reduces unnecessary steps, lowers the risk of message corruption, and keeps interactions focused and efficient, even in complex, long-horizon tasks.

Types of Agent Interactions

In contrast to previous multi-agent trading frameworks, which rely heavily on natural language dialogue, TradingAgents agents communicate primarily through structured documents and diagrams. These documents encapsulate the agents' insights in concise, well-organized reports that preserve essential content while avoiding irrelevant information. By utilizing structured reports, agents can query necessary details directly from the global state, eliminating the need for lengthy conversations that risk diluting information, extending the message state indefinitely, and causing data loss. The types of documents and the information they contain are detailed below:

• Analyst Team: Fundamental, sentiment, news, and technical analysts compile their research and findings into concise analysis reports specific to their areas of expertise. These reports include key metrics, insights, and recommendations based on their specialized analyses.
• Traders: Traders review and analyze the reports from the analysts, carefully deliberating to produce clear decision signals. They accompany these decisions with detailed reports explaining their rationale and supporting evidence, which are later utilized by the risk management team.

Agents engage in natural language dialogue exclusively during agent-to-agent conversations and debates. These concise, focused discussions have been shown to promote deeper reasoning and integrate diverse perspectives, enabling more balanced decisions in complex, long-horizon scenarios—a method particularly relevant to the intricate environment of trading. This approach seamlessly integrates with our structured framework, as the conversation state is recorded as a structured entry within the overall agent state. The types of communication in these scenarios are detailed below:

• Researcher Team: Each researcher agent queries the global agent state for analyst reports and carefully forms their opinion. Two researchers represent opposing perspectives: one bullish and one bearish. They engage in natural language dialogue for $n$ rounds, as determined by the debate facilitator agent. At the conclusion, the facilitator reviews the debate history, selects the prevailing perspective, and records it as a structured entry in the communication protocol.
• Risk Management Team: The risk management team, similar to the researcher team, queries the trader's decision and accompanying report. They then deliberate from three perspectives—risk-seeking, neutral, and risk-conservative—to adjust the trading plan within risk constraints. They engage in $n$ rounds of natural language discussion, guided by a facilitator agent.
• Fund Manager: The fund manager reviews the discussion from the risk management team, determines the appropriate risk adjustments, and updates the trader's decision and report states within the communication protocol.

Backbone LLMs

To meet the diverse complexity and speed demands of tasks in our framework, we strategically select Large Language Models (LLMs) based on their strengths. Quick-thinking models, such as gpt-4o-mini and gpt-4o, efficiently handle fast, low-depth tasks like summarization, data retrieval, and converting tabular data to text. In contrast, deep-thinking models like o1-preview excel in reasoning-intensive tasks such as decision-making, evidence-based report writing, and data analysis. These models leverage their architectures for multi-round reasoning, producing logically sound, in-depth insights. Additionally, we prioritize models with proven reliability and scalability to ensure optimal performance across various market conditions. We also employ auxiliary expert models for specialized tasks like sentiment analysis.

Specifically, all analyst nodes rely on deep-thinking models to ensure robust analysis, while quick-thinking models handle data retrieval from APIs and tools for efficiency. Researchers and traders use deep-thinking models to generate valuable insights and support well-informed decisions. By aligning the choice of LLMs with the specific requirements of each task, our framework achieves a balance between efficiency and depth of reasoning, which is crucial for effective trading strategies.

This implementation strategy ensures that TradingAgents can be deployed without requiring a GPU, relying only on API credits. It also introduces seamless exchangeability of backbone models, enabling researchers to effortlessly replace the model with any locally hosted or API-accessible alternatives in the future. This adaptability supports the integration of improved reasoning models or finance-tuned models customized for specific tasks. As a result, TradingAgents is highly scalable and future-proof, offering flexibility to accommodate any backbone model for any of its agents.

In this section, we describe the experimental setup used to evaluate our proposed framework. We also provide detailed descriptions of the evaluation metrics employed to assess performance comprehensively.

Back Trading

To simulate a realistic trading environment, we utilize a multi-asset and multi-modal financial dataset comprising of various stocks such as Apple, Nvidia, Microsoft, Meta, Google, and more. The dataset includes:

• Historical Stock Prices: Open, high, low, close, volume, and adjusted close prices from January 1st, 2024, to March 29th, 2024.
• News Articles: Daily news updates are gathered from diverse sources such as Bloomberg, Yahoo, EODHD, FinnHub, and Reddit, covering specific company developments, global events, macroeconomic trends, and government updates.
• Social Media Posts and Sentiment: Posts from Reddit, X/Twitter, and other platforms along with sentiment scores of posts calculated by auxiliary language models.
• Insider Sentiments and Transactions: Sentiment derived from public information, including transactions from SEDI and relevant company filings.
• Financial Statements and Earnings Reports: Quarterly and annual reports filed by companies.
• Company Profiles and Financial History: Descriptions of company profiles, target industries, and financial history reported by third parties.
• Technical Indicators: Sixty standard technical analysis indicators calculated for each asset, including MACD, RSI, Bollinger Bands, etc.

Simulation Setup

We simulate the trading environment for the period from June 19, 2024, to November 19, 2024. TradingAgents facilitates seamless plug-and-play strategies during the simulation, enabling straightforward comparisons with any baseline. Agents make decisions based solely on data available up to each trading day, ensuring no future data is used (eliminating look-ahead bias). Based on their analysis, TradingAgents generates trading signals to buy, sell, or hold assets, which are then executed. Afterward, analysis metrics are calculated before proceeding to the next day's data.

Baseline Models

We compare our TradingAgents framework against several baselines:

• Buy and Hold: Investing equal amounts in all selected stocks and holding them throughout the simulation period.
• MACD (Moving Average Convergence Divergence): A trend-following momentum strategy that generates buy and sell signals based on the crossover points between the MACD line and signal line.
• KDJ and RSI (Relative Strength Index): A momentum strategy combining KDJ (stochastic oscillator) and RSI (relative strength index) indicators to identify overbought and oversold conditions for trading signals.
• ZMR (Zero Mean Reversion): A mean reversion trading strategy that generates signals based on price deviations from and subsequent reversions to a zero reference line.
• SMA (Simple Moving Average): A trend-following strategy that generates trading signals based on crossovers between short-term and long-term moving averages.

Evaluation Metrics

Categories	Models	AAPL					GOOGL					AMZN
		CR%↑	ARR%↑	SR↑	MDD%↓		CR%↑	ARR%↑	SR↑	MDD%↓		CR%↑	ARR%↑	SR↑	MDD%↓
Market	B&H	-5.23	-5.09	-1.29	11.90		7.78	8.09	1.35	13.04		17.1	17.6	3.53	3.80
Rule-based	MACD	-1.49	-1.48	-0.81	4.53		6.20	6.26	2.31	1.22		-	-	-	-
	KDJ&RSI	2.05	2.07	1.64	1.09		0.4	0.4	0.02	1.58		-0.77	-0.76	-2.25	1.08
	ZMR	0.57	0.57	0.17	0.86		-0.58	0.58	2.12	2.34		-0.77	-0.77	-2.45	0.82
	SMA	-3.2	-2.97	-1.72	3.67		6.23	6.43	2.12	2.34		11.01	11.6	2.22	3.97
Ours	TradingAgents	26.62	30.5	8.21	0.91		24.36	27.58	6.39	1.69		23.21	24.90	5.60	2.11
Improvement(%)		24.57	28.43	6.57	-		16.58	19.49	4.26	-		6.10	7.30	2.07	-

Table 1: TradingAgents (AIS): Comparison of RNA Sequence (left), Modality Fusion (middle), and TradingAgents (right). Embedding base models are BERT, PubMedBERT, and OpenAI's GPT text-embedding-3-large.

Sharpe Ratio

The Sharpe Ratio performance highlights TradingAgents' exceptional ability to deliver superior risk-adjusted returns, consistently outperforming all baseline models across AAPL, GOOGL, and AMZN with Sharpe Ratios of at least 5.60—surpassing the next best models by a significant margin of at least 2.07 points. This result underscores TradingAgents' effectiveness in balancing returns against risk, a critical metric for sustainable and predictable investment growth. By excelling over market benchmarks like Buy-and-Hold and advanced strategies such as KDJRSI, SMA, MACD, and ZMR, TradingAgents demonstrates its adaptability and robustness in diverse market conditions. Its ability to maximize returns while maintaining controlled risk exposure establishes a solid foundation for multi-agent and debate-based automated trading algorithms.

Maximum Drawdown

While rule-based baselines demonstrated superior performance in controlling risk, as reflected by their maximum drawdown scores, they fell short in capturing high returns. This trade-off between risk and reward underscores TradingAgents' strength as a balanced approach. Despite higher returns being typically associated with higher risks, TradingAgents maintained a relatively low maximum drawdown compared to many baselines. Its effective risk-control mechanisms, facilitated by the debates among risk-control agents, ensured that the maximum drawdown remained within a manageable limit, not exceeding 2%. This demonstrates TradingAgents' capability to strike a robust balance between maximizing returns and managing risk effectively.

Explainability

A significant drawback of current deep learning methods for trading is their dense and complex architectures, which often render the decisions made by trading agents indecipherable to humans. This challenge, rooted in the broader issue of AI explainability, is particularly critical for trading agents, as they operate in real-world financial markets, often involving substantial sums of money where incorrect decisions can lead to severe consequences and losses.

In contrast, an LLM-based agentic framework for trading offers a transformative advantage: its operations and decisions are communicated in natural language, making them highly interpretable to humans. To illustrate this, we provide the full trading log of TradingAgents for a single day in the Appendix, showcasing its use of the ReAct-style prompting framework. Each decision made by the agents is accompanied by detailed reasoning, tool usage, and thought processes, enabling traders to easily understand and debug the system. This transparency empowers traders to fine-tune and adjust the framework to account for factors influencing decisions, offering a significant edge in explainability over traditional deep-learning trading algorithms.