Agent Market Arena (AMA)
收藏资源简介:
Agent Market Arena (AMA) 是一个用于评估基于大型语言模型 (LLM) 的交易代理的实时、终身、多类资产评估框架,完全基于经过验证且持续更新的市场数据构建。AMA 将经过专家审核的交易数据、新闻和多种代理架构集成在一个统一的交易框架中,确保在真实条件下进行公平和持续的对比。它包含四种代理,包括 InvestorAgent 作为单代理基线,TradeAgent 和 HedgeFundAgent 具有不同的风险风格,以及具有基于记忆的推理的 DeepFundAgent。AMA 在 GPT-4o、GPT-4.1、Claude-3.5-haiku、Claude-sonnet-4 和 Gemini-2.0-flash 上对这些代理进行评估。实时实验在加密货币和股票市场上都表明,代理框架展现出明显不同的行为模式,从激进的风险承担到保守的决策制定,而模型主干对结果变异的贡献较少。AMA 因此为在 LLM 基础上进行严格的、可重复的和持续发展的金融推理和交易智能评估奠定了基础。
Agent Market Arena (AMA) is a real-time, lifelong, multi-asset valuation framework for evaluating large language model (LLM)-based trading agents, built entirely on verified and continuously updated market data. AMA integrates expert-audited trading data, news, and diverse agent architectures into a unified trading framework, ensuring fair and ongoing comparison under realistic conditions. It includes four types of agents: InvestorAgent as the single-agent baseline, TradeAgent and HedgeFundAgent with distinct risk styles, and DeepFundAgent featuring memory-based reasoning. AMA evaluates these agents on GPT-4o, GPT-4.1, Claude-3.5-haiku, Claude-sonnet-4, and Gemini-2.0-flash. Real-world experiments conducted across cryptocurrency and stock markets demonstrate that the agent frameworks exhibit significantly distinct behavioral patterns, ranging from aggressive risk-taking to conservative decision-making, while the model backbone contributes relatively little to outcome variance. AMA thus lays a foundation for rigorous, reproducible, and continuously evolving evaluation of financial reasoning and trading intelligence based on LLMs.



