[논문 리뷰] TradingGPT: Multi-Agent System with Layered Memory and Distinct Characters for Enhanced Financial Trading Performance

Yang Li, Yangyang Yu외 3명|arXiv (Cornell University)|2023. 09. 07.

Stock Market Forecasting Methods인용 수 9

한 줄 요약

TradingGPT는 다중 에이전트 LLM 프레임워크로 층별 메모리와 에이전트당 거래 캐릭터를 통해 에이전트 간 토론과 다중 모달 데이터 통합으로 자동화된 금융 거래를 향상시킵니다. 메모리 계층화(단기, 중기, 장기)와 에이전트별 위험 프로필을 강조하여 거래 의사결정의 강건성 및 적응성을 높입니다.

ABSTRACT

Large Language Models (LLMs), prominently highlighted by the recent evolution in the Generative Pre-trained Transformers (GPT) series, have displayed significant prowess across various domains, such as aiding in healthcare diagnostics and curating analytical business reports. The efficacy of GPTs lies in their ability to decode human instructions, achieved through comprehensively processing historical inputs as an entirety within their memory system. Yet, the memory processing of GPTs does not precisely emulate the hierarchical nature of human memory. This can result in LLMs struggling to prioritize immediate and critical tasks efficiently. To bridge this gap, we introduce an innovative LLM multi-agent framework endowed with layered memories. We assert that this framework is well-suited for stock and fund trading, where the extraction of highly relevant insights from hierarchical financial data is imperative to inform trading decisions. Within this framework, one agent organizes memory into three distinct layers, each governed by a custom decay mechanism, aligning more closely with human cognitive processes. Agents can also engage in inter-agent debate. In financial trading contexts, LLMs serve as the decision core for trading agents, leveraging their layered memory system to integrate multi-source historical actions and market insights. This equips them to navigate financial changes, formulate strategies, and debate with peer agents about investment decisions. Another standout feature of our approach is to equip agents with individualized trading traits, enhancing memory diversity and decision robustness. These sophisticated designs boost the system's responsiveness to historical trades and real-time market signals, ensuring superior automated trading accuracy.

연구 동기 및 목표

• 금융 분야에서 인간의 인지 구조를 반영하는 메모리 중심의 다중 에이전트 거래 시스템의 필요성을 동기화한다.
• 주식 및 펀드 거래를 위한 계층적 메모리, 캐릭터 주도형 LLM 다중 에이전트 프레임워크를 제안한다.
• 다양한 관점을 활용하고 의사결정을 개선하기 위해 에이전트 간 의사소통과 토론을 가능하게 한다.
• 다중 모달 데이터 스트림과 실시간 시장 단서를 통합하여 시의적절한 거래 행동을 구현한다.
• 구성요소를 평가하고 전통적 자동 거래 방식에 비해 잠재적 성능 향상을 입증한다.]
• method
• [
• Introduce TradingGPT architecture with agents that maintain memories in three layers (short, middle, long-term) using custom decay and ranking metrics.
• Define retrieval scoring for memories using recency, relevancy, and importance with layer-specific thresholds.
• Incorporate per-agent trading characters with varying risk preferences to diversify decision-making.
• Enable inter-agent debates where agents share top memories and reflections to inform decisions.
• Use real-time multi-modal data (price, news, holdings) stored in a FAISS vector database to support cognition and retrieval.
• Outline training and testing workflows including single-agent and multi-agent phases, with prompts designed for GPT3.5 turbo and planned ablations across backbone models.]
• research_questions
• [
• How does a layered memory architecture affect the prioritization and retrieval of trading-relevant information in LLM-based agents?
• What is the impact of agent-specific trading characters on robustness and decision diversity in multi-agent financial trading?
• Can inter-agent debate and collaboration improve trading performance compared to single-agent baselines?
• How effectively can multi-modal financial data be integrated into an LLM-driven trading framework for decision making?
• What are the potential performance gains of TradingGPT relative to existing automated trading strategies under similar data conditions?]
• key_findings
• [
• The proposed layered-memory mechanism enables hierarchical retrieval of events across short-, mid-, and long-term memories.
• Inter-agent debate and diverse character profiles lead to richer insights and more robust trading decisions.
• The system integrates multi-modal data; memory-based reasoning informs daily and minute-frequency trading actions.
• Training and testing workflows combine single-agent evaluations with multi-agent debates to reflect real-world collaboration.
• The framework shows potential to outperform other automated trading strategies by emulating human cognitive dynamics and responsiveness to market changes.]
• table_headers
• []
• table_rows
• []

제안 방법

• Introduce TradingGPT architecture with agents that maintain memories in three layers (short, middle, long-term) using custom decay and ranking metrics.
• Define retrieval scoring for memories using recency, relevancy, and importance with layer-specific thresholds.
• Incorporate per-agent trading characters with varying risk preferences to diversify decision-making.
• Enable inter-agent debates where agents share top memories and reflections to inform decisions.
• Use real-time multi-modal data (price, news, holdings) stored in a FAISS vector database to support cognition and retrieval.
• Outline training and testing workflows including single-agent and multi-agent phases, with prompts designed for GPT3.5 turbo and planned ablations across backbone models.

Figure 1: TradingGPT Data Warehouse.

실험 결과

연구 질문

• RQ1 How does a layered memory architecture affect the prioritization and retrieval of trading-relevant information in LLM-based agents?
• RQ2 What is the impact of agent-specific trading characters on robustness and decision diversity in multi-agent financial trading?
• RQ3 Can inter-agent debate and collaboration improve trading performance compared to single-agent baselines?
• RQ4 How effectively can multi-modal financial data be integrated into an LLM-driven trading framework for decision making?
• RQ5 What are the potential performance gains of TradingGPT relative to existing automated trading strategies under similar data conditions?]
• RQ6 key_findings: [
• RQ7 The proposed layered-memory mechanism enables hierarchical retrieval of events across short-, mid-, and long-term memories.
• RQ8 Inter-agent debate and diverse character profiles lead to richer insights and more robust trading decisions.
• RQ9 The system integrates multi-modal data; memory-based reasoning informs daily and minute-frequency trading actions.
• RQ10 Training and testing workflows combine single-agent evaluations with multi-agent debates to reflect real-world collaboration.
• RQ11 The framework shows potential to outperform other automated trading strategies by emulating human cognitive dynamics and responsiveness to market changes.

주요 결과

• The proposed layered-memory mechanism enables hierarchical retrieval of events across short-, mid-, and long-term memories.
• Inter-agent debate and diverse character profiles lead to richer insights and more robust trading decisions.
• The system integrates multi-modal data; memory-based reasoning informs daily and minute-frequency trading actions.
• Training and testing workflows combine single-agent evaluations with multi-agent debates to reflect real-world collaboration.
• The framework shows potential to outperform other automated trading strategies by emulating human cognitive dynamics and responsiveness to market changes.

Figure 2: TradingGPT training and test workflow.