Researchers from Princeton, Ant Group and Stanford Introduce AQuA: A Two-Part Agentic Framework for Autonomous Factor Discovery and Model Development in Quantitative Finance
MarkTechPost Asif Razzaq
Princeton, Ant Group and Stanford researchers built AQuA, a two-part system for finance research. It keeps the judge fixed while the agent changes its own workflow, which is the whole point.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Quantitative research bots can fool themselves. If an agent writes the experiment, then learns from the result, a bad feature can get promoted as if it were real. The Princeton, Ant Group and Stanford team behind AQuA is trying to cut that loop off instead of polishing it with better prompts.
AQuA is split into two separate research systems. One hunts for symbolic alpha factors in crypto. The other builds time-series models for US equities. They do not share agents, memory, candidate spaces or research state, which is the whole trick: the researcher can adapt, but the evaluator does not move.
That matters because this field has a long history of convincing backtests built on small mistakes. AQuA tries to make leakage-inducing actions impossible. Each side locks in its splits, feature and label definitions, and evaluator before any iteration begins. The agent can only output a constrained factor expression or a single config change. In the authors’ words, the freedom is asymmetric.
On the factor-discovery side, AQuA uses six agents: Data Steward, Visual Analyst, Idea Miner, Factor Evaluator, Backtest Engineer and Research Librarian, all routed through an AI Manager. A factor starts as a falsifiable idea with a mechanism, a predicted direction and refutation conditions, and only then gets assembled from a standard formulaic-alpha operator set. On a crypto five-minute universe, the combined validation Spearman IC rises over 20 research epochs to about 0.190. The paper says adapted AlphaMemo reaches 0.171, AlphaGen 0.151, LSTM 0.137, LightGBM 0.106 and an Alpha158-style baseline 0.075.
The model-development side works on intraday US equities and predicts forward returns over the next thirty minutes. Training uses 2010–2019, 2020 is an embargo gap, and 2021–2025 is held out for test. Across identical data and evaluator settings, the hybrid model posts a per-stock raw IC of 0.0843, ahead of GRU at 0.0613, LSTM at 0.0535, xLSTM at 0.0434, LGB at 0.0397 and ridge at 0.0251. The paper is careful here: the two parts use different IC conventions and should not be compared directly.
The strategy layer turns the equity score into a dollar-neutral long/short book with 2 bps two-leg cost. Sector-neutralizing lifts held-out Sharpe to 2.15, a causal volatility-targeting overlay pushes it to 2.50, and a fully causal walk-forward still lands at 2.00. The backtest stays positive in every year from 2021 through 2025, including 2022, which is a nice reminder that some finance systems can survive without a séance.
My take — AI-written commentary, not fact-checked reporting
AQuA is the right kind of boring: it treats leakage like a design flaw, not a vibes problem. That should embarrass a lot of agent papers that keep bolting a reviewer on top and calling it governance. Finance has always loved a clean backtest and a dirty secret; this at least tries to lock the door before the secret walks back in.
Read more about this at: MarkTechPost