FinAgent: A multimodal foundation agent for financial trading: Data-retrieval data-mining, policy-generation, and trading

In one sentence

Summary

The paper addresses the challenge of integrating diverse, multi-modal financial data into explainable trading systems, moving beyond the limitations of traditional deep learning and reinforcement learning models that often lack interpretability and oversimplify complex market dynamics. The authors propose FinVision, a multi-agent framework designed to synthesize textual news, visual chart patterns, and historical performance data to make granular trading decisions. This approach aims to provide a transparent, step-by-step reasoning process for each trade, allowing for better risk management and strategy refinement compared to black-box models. The methodology employs four specialized agents built on GPT-4o-mini and o1-mini, orchestrated via LangGraph. The Summarize Agent processes daily news, the Technical Analyst Agent interprets candlestick charts and indicators like MACD and RSI, and the Reflection Agent analyzes past trading signals and performance metrics. These insights are fed into a Prediction Agent that outputs a BUY/SELL/HOLD action, a specific position size as a percentage of the portfolio, and a detailed rationale. The system was evaluated on Apple, Amazon, and Microsoft over a seven-month testing period in 2023, using Annual Rate of Return, Sharpe Ratio, and Maximum Drawdown as key performance indicators, while comparing against Buy-and-Hold, rule-based strategies, and RL models like PPO and DQN. Results indicate that FinVision achieved competitive returns and superior risk-adjusted performance compared to passive and RL-based baselines, with a Sharpe Ratio of 1.20 for AAPL versus 0.67 for Buy-and-Hold. Ablation studies confirmed that the reflection module significantly boosts performance by enabling adaptive learning from past errors. However, the framework underperformed the FinAgent baseline, which utilized a year-long training period, highlighting the trade-off between API costs and performance. The study demonstrates the viability of multi-modal LLM agents for trading but notes limitations in handling extreme market volatility and the high computational cost of multi-agent inference.

Research problem

Traditional financial trading models struggle with integrating multi-modal data (text, charts) and lack explainability. Existing LLM-based agents often require extensive training or lack granular risk management features like position sizing. This paper addresses the need for an explainable, multi-modal trading agent that can effectively synthesize diverse data sources to make informed, risk-aware trading decisions.

Core contributions

• Proposes FinVision, a multi-agent framework integrating news, visual technical analysis, and historical reflection for stock trading.
• Introduces a reflection module that analyzes past trading signals and performance to adapt future decisions.
• Implements granular risk management by predicting position sizes as a percentage of the portfolio.
• Demonstrates superior risk-adjusted returns compared to traditional rule-based and reinforcement learning baselines.
• Provides detailed explainability for trading decisions through Chain of Thought reasoning in each agent.

Method

• Summarize Agent: Condenses daily news into factual summaries using GPT-4o-mini.
• Technical Analyst Agent: Analyzes candlestick charts and technical indicators using vision capabilities.
• Reflection Agent: Assesses short/medium-term performance and visualizes past trading signals for feedback.
• Prediction Agent: Integrates all insights to output action, position size, and explanation using o1-mini.
• LangGraph orchestration: Manages state and dependencies between specialized agents.
• Risk management rules: Enforces cash reserves and position sizing constraints in the decision logic.

Results & findings

• FinVision achieved a 14.79% ARR and 1.20 Sharpe Ratio for AAPL, outperforming Buy-and-Hold (13.56% ARR, 0.67 SR).
• For MSFT, FinVision reached 25.57% ARR and 1.41 SR, surpassing the market's 22.27% ARR and 1.01 SR.
• AMZN results showed 42.14% ARR with a significantly lower MDD (12.09%) compared to Buy-and-Hold (17.45%).
• RL models (PPO, DQN) performed poorly with negative Sharpe Ratios across all tested stocks.
• Ablation studies showed that removing the reflection module significantly degraded performance.
• The framework underperformed FinAgent, which used a year-long training period, indicating a data-efficiency trade-off.

Data & experiments

• Evaluated on Apple (AAPL), Amazon (AMZN), and Microsoft (MSFT) from April to December 2023.
• Data sources include Yahoo Finance news, daily candlestick charts, and technical indicators (MACD, RSI, Bollinger Bands).
• Two-month training period for reflection data, followed by a seven-month testing period.
• Benchmarks include Buy-and-Hold, MACD, RSI, PPO, DQN, and the FinAgent model.
• Metrics used: Annual Rate of Return (ARR), Sharpe Ratio, and Maximum Drawdown (MDD).

Limitations & caveats

• Underperformed FinAgent, which used a year-long training period, suggesting data volume is critical.
• Evaluated only on three large-cap tech stocks, limiting generalizability to other sectors.
• High API costs associated with multi-agent inference and vision processing.
• Reflection module relies on historical data, which may not predict future black swan events.
• Tested in a bullish market, which may favor certain strategies over others.

Curator notes

Discovery status: institutional_trading_round2_promoted. Source links: A Multi-Agent Orchestration Framework for Venture Capital Due Diligence