跳转至

文章背景与核心概要

自动化Alpha因子挖掘旨在严格的评估约束下,从价量面板和订单簿数据中发现符号化交易信号。传统的单智能体和多智能体系统经常面临两大主要缺陷:一是过度拟合预测代理指标,在考虑执行成本后往往失效;二是冗余地探索熟悉的因子家族,限制了行为多样性和运营鲁棒性。

为了解决这些局限性,本文推出了 GoAnt,这是一个质量-多样性(Quality-Diversity)多智能体搜索框架。GoAnt 通过共享的自适应心智图(Mental Map)和一个精简的女王调度器(Queen dispatcher),协调具有专业分工且不进行直接通信的工作者角色(探索者 Explorer、利用者 Exploiter 和连接者 Connector)。


GoAnt: Quality-Diversity Multi-Agent Search for Alpha Factor Discovery in Market Microstructure Data

Authors: Stella Zhao, Tommy Sha
ArXiv ID: arXiv:2609.08719 [cs.AI]
Submitted: September 8, 2026
Subjects: Artificial Intelligence (cs.AI)


📌 Summary

自动化alpha因子挖掘旨在从价量面板和订单簿数据中,在严格的评估约束下发掘符号化交易信号。传统的单智能体和多智能体系统经常遭受两大主要缺陷:过度拟合预测代理指标(在计入执行成本后便会崩溃),以及冗余地探索熟悉的因子家族(这限制了行为多样性和运营鲁棒性)。

Automated alpha factor discovery aims to uncover symbolic trading signals from price-volume panels and order-book data within strict evaluation constraints. Traditional single- and multi-agent systems frequently suffer from two major flaws: overfitting predictive proxies that collapse after accounting for execution costs, and redundantly exploring familiar factor families, which limits behavioral diversity and operational robustness.

为了解决这些局限性,本文引入了 GoAnt,一个质量-多样性多智能体搜索框架。GoAnt 通过共享的自适应心智图和一个精简的女王调度器,协同各个专业化且不通信的工作者角色(探索者、利用者和连接者)。

To resolve these limitations, this paper introduces GoAnt, a quality-diversity multi-agent search framework. GoAnt coordinates specialized, non-communicating worker roles (Explorer, Exploiter, and Connector) through a shared, adaptive Mental Map and a compact Queen dispatcher.

核心创新点:

Key Innovations:

  • 心智图与女王调度器: 候选因子按照无信息泄露的执行配置文件进行系统化组织,每个生态位(niche)保留一个精英候选因子。女王通过显式的搜索状态摘要动态重新分配评估预算。
  • The Mental Map & Queen Dispatcher: Candidates are systematically organized by leakage-free execution profiles, retaining one elite candidate per niche. The Queen dynamically reallocates evaluation budgets using explicit search-state summaries.
  • 有效收益协议(Effective-Yield Protocol): 一种与地图无关的评估协议,直接从原始评估日志中统计高质量、互不冗余的因子,为基于存档和无地图的方法提供了一个统一的基准。
  • Effective-Yield Protocol: A map-independent evaluation protocol that directly counts high-quality, mutually non-redundant factors from raw evaluation logs, offering a unified benchmark for both archive-based and map-free methods.
  • 卓越性能: 在真实世界的A股微观结构数据(2023–2026年)上进行评估,在同等预算下,GoAnt的质量加权收益达到了 41.8(价量)和 47.6(订单簿),分别比最强的基线高出 57%97%。此外,与静态地图架构的 0.61 和 0.63 相比,GoAnt 锁定的群体保留了更强的样本外质量(达到样本内表现的 0.64 和 0.67)。
  • Superior Performance: Evaluated on real-world A-share microstructure data (2023–2026), GoAnt achieves quality-weighted yields of 41.8 (price-volume) and 47.6 (order-book), outperforming the strongest baseline by 57% and 97% under matched budgets. Furthermore, GoAnt's locked populations retain robust out-of-sample quality (0.64 and 0.67 of in-sample performance) compared to 0.61 and 0.63 for static map architectures.

📋 Bibliographic & Access Information