Gate研究院:多智能体LLM交易框架在BTC回测中显著跑赢Buy & Hold策略

Odaily星球日报讯 Gate 研究院近日发布的《基于 Multi-Agent LLM 的 BTC 交易框架研究与回测分析》指出,相比单一 LLM 直接生成交易信号,多智能体(Multi-Agent)LLM 架构更接近真实金融机构的投研流程,能够通过分析师、研究员、交易员与风控团队之间的协作与辩论,提升交易决策的透明度与风险控制能力。研究基于 TradingAgents 框架,对 BTC 市场构建了适用于加密场景的 AI 交易系统,并引入技术分析、新闻分析、情绪分析与宏观/链上分析等多类智能体角色。
研究基于 BTC/USDT 1 小时级别数据,对 TradingAgents-BTC 策略进行了历史回测。结果显示,该策略在测试区间内取得 +20.25% 总收益,显著优于同期 Buy & Hold 的 -7.89%,同时最大回撤控制在 -17.41%,低于 Buy & Hold 的 -27.06%。研究认为,多智能体框架在震荡与下跌阶段能够通过 Sell/Underweight 与 Flat 状态减少部分风险暴露,并在市场反弹阶段重新切入多头,从而提升整体风险收益表现。
报告指出,Multi-Agent LLM 框架在加密交易场景中展现出一定应用潜力,但当前回测周期仅覆盖约三个月,且 1 小时级别交易仍可能受到手续费、滑点与信号延迟影响。未来仍需在更长历史周期、不同市场环境与更多资产类别中进一步验证策略稳定性与泛化能力。
Disclaimer: OKX Orbit content is provided for informational purposes only. Learn more
Replies
Related Flash News
Analysis: The continued rise in U.S. Treasury yields has weakened market willingness to allocate to Bitcoin
Nearly $320 million in Bitcoin long positions were liquidated due to the SEC's delayed plan
CFTC Chair: No longer allow government involvement in cryptocurrencies
Analysis: $1.26 billion Bitcoin ETF outflow may become a 'reverse buying signal'
About 1,400 electricity theft Bitcoin mining machines were seized in São Paulo state, Brazil
Since its launch in 2020, JPMorgan's Kinexys platform has seen over $1.5 trillion in transaction volume
Data: Morgan Stanley increased its holdings by another 14 Bitcoins, bringing total holdings to 3,486
Bitdeer currently maintains zero interest, having sold 201.6 BTC this week
Garrett Jin short over 57,000 ZEC on 3x terms, with a 5x BTC long position showing a floating loss of $920,000
Michael Saylor: Strategy does not rule out selling some Bitcoin in 2026



