普林斯顿、蚂蚁集团与斯坦福团队提出 AQuA：面向量化金融的两段式智能体自动因子发现框架

定量研究代理如果自己编写实验，可能会破坏它们随后从中学习的证据。一个表现良好的泄露特征会被存储为成功的先例，并在后续迭代中传播。提示级指令和评审代理无法解决这一问题，因为作者和评审者具有相同的盲点。来自普林斯顿大学、蚂蚁集团和斯坦福大学的一个研究团队提出了AQuA：https://arxiv.org/abs/2608.12841。AQuA是一对由语言模型驱动的研究系统，它们在迭代中改进自己的研究过程，而评估它们的系统保持不变。一个发现加密货币的符号alpha因子；另一个开发美国股票的时间序列模型。它们不共享代理、记忆、候选空间或研究状态。

定量研究在小的技术错误上会崩溃，这些错误会生成令人信服但不可重复的回测，Bailey等人已有记录：https://www.ams.org/journals/notices/201405/rnoti-p458.pdf。一个自己编写实验的代理会使情况更糟：表现良好的泄露特征会被存储为先例，而递归会像真实发现一样放大未检测到的错误。

提示级指令和模型评审不是完整的诚信边界。反复访问固定的保留集会引起自适应过拟合：https://proceedings.mlr.press/v37/blum15.html，并且已观察到大语言模型代理利用错误设定的目标和评估器：https://arxiv.org/abs/2503.11926。AQuA则使引发泄露的行为不可用。每个部分在任何迭代开始前都会固定其数据分割、特征和标签定义以及评估器，代理仅输出受约束的因子表达或单一配置差异。研究团队称之为不对称自由：代理在其DSL内部自由探索，而评估器则位于自适应表面之外。改进的对象是研究过程。

第一部分是一个六代理管道：数据管理员、可视化分析师、想法挖掘者、因子评估员、回测工程师和研究图书管理员——由AI经理协调。代理之间从不直接调用；每次交接都通过经理进行，使运行过程可审计。

因素作为可证伪的提议进入，而非表达式：假设、机制、预测方向和反驳条件。只有在这时，才从标准的公式-α算子注册表中组装出来：https：//arxiv.org/abs/1601.00991。由于每个时间序列算子只读取尾随窗口，每个截面算符只读取当前时间戳，因果关系在合成下是封闭的。运行三个反馈循环：回测中的方向校准，运行中的反证驱动信念更新，以及引导下一次搜索的交叉运行记忆。

在加密五分钟宇宙中，Spearman IC的综合验证跨越20个研究纪元，达到约0.190，而改良版AlphaMemo为0.171：https：//arxiv.org/abs/2606.20625，改良版AlphaGen为0.151：https：//arxiv.org/abs/2306.12964，LSTM为0.137，LightGBM为0.106，Alpha158风格基线为0.075。各机制保持弱势——单因子IC为0.026至0.037。该声明关注的是线束，而非单一表达式。

第二部分预测每支股票在未来三十分钟内对美国盘中股票的远期回报。2010–2019年的培训运行，2020年为禁运缺口，未被触及，2021–2025年为未动测试数据。选择仅使用培训窗口末的内部验证切片。

这里的假设是一个配置差异——架构、损耗、采样器或优化器——而一个配置只产生一个变体，保持变体的可比较性。预测变量是混合型的：多尺度一维卷积前端，可配置的跨LSTM主干网，Mamba：https：//arxiv.org/abs/2312.00752，注意点：https：//arxiv.org/abs/1706.03762（报告运行中注意），一个跨面板混合的横截面级，门控融合和按库存数据汇集。

没有单一的价格-交易量特征携带信号：最强的是5分钟收益为−0.031，而岭回归组合仅达到+0.025。在相同数据和相同评估器下，不同模型家族的每股原始IC分别为+0.0251（岭回归）、+0.0397（LGB）、+0.0434（xLSTM：https://arxiv.org/abs/2405.04517）、+0.0535（LSTM）、+0.0613（GRU），混合模型为+0.0843——比最佳基线绝对提高+0.0230，相对提高37.5%。两部分的IC使用不同的标准，论文明确表示不应进行比较。

每股评分变为一个美元中性阈值的多空组合，双向成本为2个基点。进行行业中性化将持出Sharpe提高到+2.15，训练值和持出值几乎相等。因果波动率目标覆盖可将其提高至+2.50，而完全因果的滚动前瞻模型（只使用过去数据选择每个参数）仍可达到+2.00。每股R²为1.20%。按年份计算的Sharpe分别为2021至2025年的+1.7、+3.5、+1.9、+1.8和+2.7——每年均为正值，包括2022年的下滑期。

需要与我们合作推广您的GitHub仓库、Hugging Face页面、产品发布或网络研讨会等吗？请联系：https://forms.gle/wbash1wF6efRj8G58

注意：感谢蚂蚁研究团队对本文提供的思想领导力/资源支持。蚂蚁研究团队为本文/内容的推广提供了支持。

Asif Razzaq是Marktechpost Media Inc.的首席执行官。作为一名有远见的企业家和工程师，Asif致力于利用人工智能的潜力造福社会。他最近的项目是推出人工智能媒体平台Marktechpost，该平台因其对机器学习和深度学习新闻的深入报道而脱颖而出，内容既具有技术深度又易于广大受众理解。该平台每月访问量超过200万次，显示出其在受众中的受欢迎程度。

Quantitative research agents that write their own experiments can corrupt the evidence they later learn from. A leaky feature that scores well gets stored as a successful precedent and propagated through later iterations. Prompt-level instructions and reviewer agents do not close this, because author and reviewer share the same blind spots. A team of researchers from Princeton University, Ant Group and Stanford University propose AQuA ：https://arxiv.org/abs/2608.12841 . AQuA is a pair of language-model-driven research systems that improve their own research process across iterations while the thing judging them stays frozen. One discovers symbolic alpha factors on crypto; the other develops time-series models on US equities. They share no agents, memories, candidate spaces or research state.

Quantitative research breaks on small methodological errors that produce convincing but non-reproducible backtests, documented since Bailey et al.：https://www.ams.org/journals/notices/201405/rnoti-p458.pdf. An agent writing its own experiments makes this worse: a leaky feature that scores well gets stored as precedent, and recursion amplifies an undetected bug as readily as a real discovery.

Prompt-level instructions and model review are not an integrity boundary. Repeated access to a fixed holdout causes adaptive overfitting：https://proceedings.mlr.press/v37/blum15.html, and LLM agents have been observed exploiting misspecified objectives and evaluators：https://arxiv.org/abs/2503.11926. AQuA instead makes leakage-inducing actions unavailable. Each part fixes its splits, feature and label definitions and evaluator before any iteration starts, and the agent emits only a constrained factor expression or a single config diff. The research team call this asymmetric freedom : the agent explores freely inside its DSL, but the evaluator sits outside the adaptive surface. What improves is the research process.

Part I is a six-agent pipeline: Data Steward, Visual Analyst, Idea Miner, Factor Evaluator, Backtest Engineer and Research Librarian — orchestrated by an AI Manager. Agents never call one another; every handoff goes through the Manager, keeping runs auditable.

A factor enters as a falsifiable proposal, not an expression: hypothesis, mechanism, predicted direction, and refutation conditions. Only then is it assembled from the standard formulaic-alpha operator registry：https://arxiv.org/abs/1601.00991. Because every time-series operator reads only a trailing window and every cross-sectional operator reads only the current timestamp, causality is closed under composition. Three feedback loops run: direction calibration inside a backtest, falsification-driven belief update inside a run, and cross-run memory that steers the next search.

On a crypto five-minute universe the combined validation Spearman IC climbs across 20 research epochs to approximately 0.190 , against 0.171 for an adapted AlphaMemo：https://arxiv.org/abs/2606.20625, 0.151 for an adapted AlphaGen：https://arxiv.org/abs/2306.12964, 0.137 for LSTM, 0.106 for LightGBM and 0.075 for an Alpha158-style baseline. Individual mechanisms stay weak — single-factor ICs of 0.026 to 0.037. The claim is about the harness, not one expression.

Part II predicts each stock’s forward return over the next thirty minutes on intraday US equities. Training runs on 2010–2019, 2020 is an embargo gap nothing touches, and 2021–2025 is untouched test data. Selection uses an inner-validation slice from the end of the training window only.

A hypothesis here is one config diff — architecture, loss, sampler or optimizer — and one diff produces exactly one variant, keeping variants comparable. The predictor is a hybrid: a multi-scale 1-D convolutional front-end, a configurable backbone spanning LSTM, Mamba：https://arxiv.org/abs/2312.00752 and attention：https://arxiv.org/abs/1706.03762 (attention in the reported run), a cross-sectional stage that mixes across the panel, gated fusion and a pooled per-stock readout.

No single price-volume feature carries the signal: the strongest is a 5-minute return at −0.031, and a ridge combination reaches only +0.025. Across model families on identical data and the same evaluator, per-stock raw IC runs +0.0251 (ridge), +0.0397 (LGB), +0.0434 (xLSTM：https://arxiv.org/abs/2405.04517), +0.0535 (LSTM), +0.0613 (GRU) and +0.0843 for the hybrid — +0.0230 absolute over the best baseline, 37.5% relative. The two parts’ ICs use different conventions and the paper states plainly they should not be compared.

The per-stock score becomes a dollar-neutral threshold long/short book at a two-leg cost of 2 bps. Sector-neutralizing raises the held-out Sharpe to +2.15 , with training and held-out values nearly equal. A causal volatility-targeting overlay lifts it to +2.50 , and a fully causal walk-forward choosing every parameter from past data alone still reaches +2.00 . Per-stock R² is 1.20%. Sharpe by year runs +1.7, +3.5, +1.9, +1.8 and +2.7 for 2021 through 2025 — positive in every year, including the 2022 drawdown.

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us ：https://forms.gle/wbash1wF6efRj8G58

Note:Thanks to the Ant Research team for the thought leadership/ Resources for this article. Ant Research team has supported this content/article for promotion.

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.