AQuA框架提出量化AI研究护栏，聚焦因子发现与模型评测 AQuA Proposes Guardrails for AI Quant Research, From Factor Discovery to Model Evaluation
来自普林斯顿大学、蚂蚁集团和斯坦福大学的研究团队提出AQuA，以固定数据切分、标签定义和评测器的方式约束AI代理实验流程。论文报告，该方法在加密货币因子发现及美股短线收益预测任务中优于所比较的基线方案。 Researchers from Princeton University, Ant Group, and Stanford University introduced AQuA, an agentic framework that fixes data splits, label definitions, and evaluators while allowing constrained AI experimentation. The paper reports improvements over compared baselines in crypto factor discovery and U.S. equity short
量化研究中的AI代理可以自动提出假设、编写实验并解读回测，但这也可能放大数据泄漏、过拟合或评测偏差。一旦存在缺陷的因子被记录为“成功经验”，后续代理可能继续沿用并强化这一错误。AQuA的目标是把研究探索与评价规则分开，降低自动化流程自我确认的风险。
AQuA如何限制代理的自由度
AQuA采用“非对称自由”设计：代理可在限定的表达空间中探索研究方案，但数据分割、特征与标签定义、以及评测器在实验开始前固定，不能随着结果变化。研究者认为，这种安排让流程可以迭代，同时保留可比较、可追溯的判分标准。
在因子发现部分，系统由数据、可视化、创意挖掘、因子评估、回测和资料整理等六类代理组成，并由AI Manager协调。代理不直接相互调用，交接经由Manager完成。每个候选因子先被表达为可证伪假设，说明其机制、预期方向及可能推翻该假设的条件，再转化为公式化alpha算子。
论文报告的实验结果
研究重点不是寻找某个“神奇”交易公式，而是让AI代理在固定、可审计的研究边界内开展实验。
AI agents can generate hypotheses, write experiments, and interpret backtests for quantitative research, but they can also amplify leakage, overfitting, and evaluation bias. If a flawed factor is retained as a successful lesson, later agents may reuse and reinforce it. AQuA aims to separate research exploration from evaluation rules to reduce this form of automated self-confirmation.
How AQuA Constrains Agent Freedom
AQuA uses an “asymmetric freedom” design. Agents may explore within a bounded representation space, while data splits, feature and label definitions, and evaluators are fixed before experimentation and do not change with results. The researchers argue that this permits iteration while preserving comparable and traceable scoring rules.
For factor discovery, the system includes six agent roles covering data, visualization, idea mining, factor evaluation, backtesting, and documentation, coordinated by an AI Manager. Agents do not call one another directly; handoffs pass through the Manager. Each candidate factor is first expressed as a falsifiable hypothesis with a mechanism, expected direction, and disconfirming conditions before becoming a formulaic alpha operator.
Reported Experimental Results
The focus is not a single “magic” trading formula, but allowing AI agents to experiment within fixed, auditable research boundaries.
来源
- MarkTechPost · 09-01 23:54