跳转至

文章背景与核心概要

近年来，大语言模型（LLM）在金融科技与量化投资领域的应用取得了显著进展，但如何实现真正自主、具备持续进化能力的研究体系依然是一个巨大的挑战。本文介绍的 AQuA 框架旨在研究量化投资研究中的“递归自我提升”（recursive self-improvement），核心探讨自主 AI 系统是否能够利用前序实验的实证证据来改进后续的假设和候选模型。

AQuA 框架由两个独立运行且不共享记忆、候选空间、智能体或研究状态的语言模型驱动系统组成：一是符号化因子发现系统，二是可训练模型开发系统。两个系统均在严格控制的封闭沙箱环境中运行，通过保留经验证的证据来引导未来的迭代。实验表明，该框架在加密货币和美股市场均展现出极高的信息系数（IC）以及在扣除交易成本后高达 \(+2.50\) 的夏普比率。

AQuA: Recursively Self-Improving Quantitative Trading Research Agents

Authors: Jiacheng Guo, Suozhi Huang, Yunlong Gao, Zihao Li, Jian Ge, Xu Kuang, Mengdi Wang
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv ID: arXiv:2608.12841
Submitted: August 13, 2026

Executive Summary

AQuA 是一个专为研究量化投资研究中的递归自我提升（recursive self-improvement）而设计的自主研究框架。其核心前提是探讨自主 AI 系统是否能够利用前序实验的实证证据来提炼后续的假设和候选模型。

AQuA is an autonomous research framework designed to study recursive self-improvement in quantitative investment research. The core premise investigates whether an autonomous AI system can utilize empirical evidence from prior experiments to refine subsequent hypotheses and candidate models.

该框架由两个独立的、由语言模型驱动的研究系统组成，它们不共享记忆、候选空间、智能体或研究状态： 1. 符号化因子发现系统： 一个由管理者协调的多智能体流水线，专门负责发现和组合市场因子。 2. 可训练模型开发系统： 一个围绕混合时间序列架构构建的、由配置驱动的循环。

每个系统都在严格控制的封闭沙箱环境中，通过保留经过验证的证据来指导未来的迭代，从而独立闭环其研究反馈循环。

The framework consists of two independent language-model-driven research systems that do not share memories, candidate spaces, agents, or research states: 1. Symbolic Factor Discovery System: A manager-mediated multi-agent pipeline dedicated to discovering and combining market factors. 2. Trainable Model Development System: A configuration-driven loop built around a hybrid time-series architecture.

Each system independently closes its own research feedback loop by retaining validated evidence to guide future iterations within a strictly controlled, sealed sandbox environment.

Key Methodology & Architecture

  • 封闭沙箱设计： 两个研究循环均在隔离的环境中运行，固定了数据划分、特征/标签定义以及评估器。语言模型被严格限制只能通过因子表达式或配置差异（diffs）进行交互。
  • 独立的优化流水线： 通过将因子系统与模型系统隔离，AQuA 防止了研究状态的交叉污染，从而能够针对每种特定的范式进行有针对性的递归自我提升。
  • Sealed Sandbox Design: Both research loops operate inside isolated environments that fix data splits, feature/label definitions, and evaluators. The language models are strictly constrained to interacting via factor expressions or configuration diffs.
  • Independent Optimization Pipelines: By segregating the factor and model systems, AQuA prevents cross-contamination of research states, enabling targeted recursive self-improvement tailored to each specific paradigm.

Performance Highlights

  • 符号化因子发现： 成功将发现的因子组合成稳健的信号，在加密货币市场生态中实现了约 \(0.190\) 的组合信息系数（IC）。
  • 可训练模型开发： 在美股市场上实现了每只股票 \(+0.0843\) 的 IC，并成功转化为阈值多空策略。
  • 风险调整后收益： 实现了高达 \(+2.50\) 的测试集夏普比率（考虑了两条腿的交易成本），并且在 2021年至2025年的每一年中均保持正收益。
  • Symbolic Factor Discovery: Successfully combines discovered factors into a robust signal, achieving a combined Information Coefficient (IC) of approximately \(0.190\) on a cryptocurrency market universe.
  • Trainable Model Development: Achieves a per-stock IC of \(+0.0843\) on US equities, which is effectively translated into a threshold long/short strategy.
  • Risk-Adjusted Returns: Delivers a held-out Sharpe ratio of up to \(+2.50\) (accounting for a two-leg transaction cost) and maintains positive performance every single year from 2021 to 2025.