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

**Authors:** Yang Li, Yangyang Yu, Haohang Li, Zhi Chen, Khaldoun Khashanah  
**arXiv:** [2309.03736](https://arxiv.org/abs/2309.03736)  
**Date:** 2023-09-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.03736v1/

## Key points

- The demand for intelligent trading robot-agents is increasing in the financial technology landscape.
- Researchers are exploring innovative approaches to design advanced trading systems using technologies like Large Language Models (LLMs) and deep learning algorithms.
- LLMs, such as the Generative Pre-trained Transformers (GPT) series, have revolutionized artificial intelligence and natural language processing across various domains.
- A novel LLM-powered multi-agent framework with layered memories has been introduced for stock and fund trading applications to address challenges in memory processing mechanisms.
- The framework incorporates a collaborative approach where agents organize memory into distinct layers governed by custom decay mechanisms, enhancing decision-making processes through inter-agent debates and individualized trading traits integration.
- Recent advancements in generative agent systems with memory streams and customized character designs have expanded LLM capabilities for multi-agent systems, enabling dynamic interactions and improved mathematical reasoning accuracy.
- TradingGPT's development involves integrating multi-modal financial data from various sources to formulate multi-layer memories for agents and train them using historical trading records.

## Summary

In the rapidly evolving landscape of financial technology, the demand for intelligent trading robot-agents is on the rise. These agents can effectively synthesize and interpret diverse data sources to inform investment decisions. Researchers are continuously exploring innovative approaches to design advanced trading systems that leverage cutting-edge technologies such as Large Language Models (LLMs) and deep learning algorithms. The emergence of LLMs, particularly exemplified by the Generative Pre-trained Transformers (GPT) series, has revolutionized artificial intelligence and natural language processing. These models have demonstrated remarkable proficiency across various domains, from healthcare diagnostics to analytical business reports. However, the existing memory processing mechanisms of LLMs do not precisely mirror the hierarchical nature of human memory, leading to challenges in efficiently prioritizing critical tasks. To address this gap, a novel LLM-powered multi-agent framework with layered memories has been introduced for stock and fund trading applications. This framework incorporates a collaborative approach where agents organize memory into distinct layers governed by custom decay mechanisms, aligning more closely with human cognitive processes. Additionally, agents can engage in inter-agent debates to enhance decision-making processes. One of the key features of this innovative framework is the integration of individualized trading traits among agents, which enhances memory diversity and decision robustness. By leveraging a collaborative multi-agent system and capturing market dynamics from varied perspectives, this approach significantly improves automated trading outcomes by fostering interactions among agents and capturing intricate market dynamics. Furthermore, recent advancements in generative agent systems with memory streams and customized character designs have expanded LLM capabilities for multi-agent systems. These systems exhibit human-like behaviors while retaining individual characters, enabling dynamic interactions with peers and environments to forge memories and relationships. Additionally, a debate mechanism leveraging multiple language models in a multi-agent system has been introduced to enhance mathematical reasoning and factual accuracy in generated content. This mechanism allows various model instances to propose debates and collaboratively converge on unified answers. In terms of data integration for TradingGPT's development, an extensive array of multi-modal financial data has been systematically integrated from various sources such as financial databases and APIs. This data serves dual purposes: formulating multi-layer memories for agents and training them using historical trading records to refine their decision-making processes. Overall, these sophisticated designs in financial AI incorporating Multi-Modal Learning techniques are poised to revolutionize automated trading strategies by enhancing collaboration among agents, capturing diverse market perspectives, and improving decision-making processes based on comprehensive historical insights.