PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yang Li, Yangyang Yu, Haohang Li, Zhi Chen, Khaldoun Khashanah

arXiv:2309.03736v1q-fin.PMq-fin.TR

TL;DR

Financial LLM trading systems can rely on single-agent, text-only inputs or treat memory as an undifferentiated whole, while earlier trading agents have limitations in representing memory and social interaction. TradingGPT combines layered memory, character profiles, multimodal financial data, and agent debate; the authors state that its layered-memory approach improves robustness.

  • Problem

    Prior multi-agent trading methods are limited in representing agents’ memory and evolving social interaction, while financial LLM systems often use single-agent text-only setups or process memory as a whole.

    Source: §1 · p. 2

  • Method

    TradingGPT combines LLM agents with short-, middle-, and long-term memory, distinct trading characters, multimodal financial data, and inter-agent debate.

    Source: §1 · p. 2

  • Results

    The authors state that the layered-memory approach improves system robustness.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors aim for an LLM-based trading system adaptable to different LLMs and identify possible applications in other sectors, including business, healthcare, and technology.

    Source: §5 · p. 6

Abstract

from arXiv · show

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.

1 Introduction

TradingGPT is an LLM-powered multi-agent trading framework that combines layered memories, debate, and customized characters. It is designed to address limits in trading agents’ memory and social interaction while incorporating diverse financial data.

1 Introduction

  • Earlier multi-agent trading methods were described as limited in modeling memory archives and evolving social interaction, while financial LLM systems often used single-agent, text-only setups or processed memory as a whole.The authors identify prioritizing immediate and critical memory events as a challenge for those systems.Source: §1 · p. 2
  • The paper presents TradingGPT as a framework for automated trading and optimal execution, and says it methodically evaluates system components.The introduction describes outperforming other automated strategies at representative fund firms such as ARK as a prospect, not a reported result.Source: §1 · p. 2
  • TradingGPT organizes agents’ memories into short-, middle-, and long-term layers and combines memory streams with debate to support reflection on past and present events and collective decision-making.The authors state that this approach improves system robustness.Source: §1 · p. 2
  • TradingGPT assigns agents different risk preferences and industry investment scopes, with the stated potential to reflect human intuition and uncover latent market opportunities.The described profiles include risk-seeking, risk-neutral, and risk-averse agents.Source: §1 · p. 2
  • The system integrates real-time multimodal information across macro and micro perspectives with historical trading records, using daily and minute-by-minute updates.The authors say these update frequencies support prompt daily-trade responses and offer high-frequency trading capability.Source: §1 · p. 2

2 Related Work

Related work covers advances in LLMs, generative agents with memory and character design, and debate among language-model agents.

2.1 Large language models (LLMs)

  • The section traces language models from Word2Vec and GloVe through BERT to GPT and Llama models, noting the latter’s proficiency across applications.The cited markers are Word2Vec [4], GloVe, BERT [2], GPT [9], and Llamas [15].Source: §2.1 · p. 2

2.2 Generative agent system with memory streams and customized character design

  • Generative agent work combines separate memory streams and character profiles, with agents forming memories and relationships and coordinating tasks through natural language.Park et al. [10] is identified as introducing memory streams and character design in this line of work.Source: §2.2 · p. 3

2.3 Multi-agent debate mechanism

  • A multi-agent debate mechanism lets language-model instances propose and discuss answers before converging on a unified answer.Du et al. [3] report benefits for mathematical and strategic reasoning and factual accuracy.Source: §2.3 · p. 3

3 Dataset and Database Structure

The dataset and database structure section describes multimodal financial data used to form agent memories and train, guide, and back-test agents against ARK historical trades.

3 Dataset and Database Structure

  • The study integrates financial data from August 15, 2020, to August 15, 2023, from sources including Databento, Alpaca News, and ARK holdings records.The passage also mentions audio transcriptions from ARK Invest videos translated to text with Whisper API.Source: §3 · p. 3
  • The data is used both to formulate agents’ multilayer memories and to train, guide, and back-test them using ARK funds’ historical trading records.Source: §3 · p. 3
  • The authors use FAISS for vector storage and semantic searches, citing fast querying of their data and compatibility with OpenAI and cosine-similarity computation for specific tickers.Source: §3 · p. 3

4 Proposed Method

TradingGPT integrates LLM-driven trading agents that manage layered memories, individual reflections, and inter-agent deliberation. Its training and testing workflows use these elements to inform trading decisions.

4.1.1 Layered-memory structure

  • Agents sort memories into short-, middle-, and long-term layers, then rank events within each layer by recency, relevancy, and importance.Layer-specific thresholds and an add-counter promote frequently triggered, profit- or loss-associated events toward longer-term retention.Source: §4.1.1 · p. 4
  • Recency decays with elapsed time at a rate controlled by layer stability, while relevancy measures cosine similarity between event and prompt embeddings.The three normalized scores are linearly combined for ranking; the reported layer thresholds are 80, 60, and 40, and scores below 20 are removed.Source: §4.1.1 · p. 4

4.1.2 Memory formulated by individual experience

  • Macro-level indicators enter long-term memory, quarterly strategies middle-term memory, and daily investment messages short-term memory.Agents draw on these memories in single-agent and multi-agent workflows.Source: §4.1.2 · pp. 4–5
  • Daily immediate reflections combine retrieved events and market facts into five trade recommendations with justifications, trade values, volumes, and returns.Extended reflections provide a broader, period-based performance overview and serve as a supplementary reference for recent transactions.Source: §4.1.2 · p. 5

4.1.3 Memory gained by interacting with other agents

  • For stocks held by multiple agents, debate shares top-K layered memories and immediate reflections, with feedback stored in the recipient’s debate memory.Shared material includes recommendations, trade values, volumes, and returns.Source: §4.1.3 · p. 5

4.2.1 Training

  • Training first generates single-agent signals from top-K memories and market data, then adds same-stock debates among agents from different sectors.The training workflow uses daily fund records and stock closing prices to support immediate reflections and trading actions.Source: §4.2.1 · pp. 5–6

4.2.2 Test

  • Testing combines individual memories and inter-agent insights, but omits representative-fund trading records and uses daily stock prices as market facts.A week of prior training reflections and debates supplies auxiliary references when market ground truth is limited.Source: §4.2.2 · p. 6

5 Current Stage And Future work

The section describes prompt design and planned ablation studies, then states TradingGPT’s performance goal and possible applications.

5 Current Stage And Future work

  • The study describes efficient prompts built with GPT3.5 turbo as the backbone for the training and testing workflow.Figure 3 is identified as showing examples of the prompt design.Source: §5 · p. 6
  • Planned ablations compare GPT3.5 turbo and CodeLlama 34B with multi-agent reinforcement learning across specified training and testing periods.The proposed metrics are cumulative trade returns, volatility, and the Sharpe Ratio.Source: §5 · p. 6
  • The stated goal is an LLM-based multi-agent trading system with layered memory and character design that aims to outperform other leading trading-agent systems.The authors also propose potential applications in games and robo-consulting across business, healthcare, and technology.Source: §5 · p. 6

Related papers

Loading 2309.03736v1…