Researchers from Princeton, Ant Group, and Stanford Propose AQuA Framework for Leakage-Free Quant Finance Research

Quantitative research that lets agents write their own experiments can taint the evidence they later learn from. A leaky feature that scores well becomes a stored precedent and propagates through iterations. Prompt-level instructions and reviewer agents do not close this gap because authors and reviewers share the same blind spots. AQuA addresses this by making leakage-inducing actions unavailable.

AQuA consists of two independent language-model-driven research systems: one discovers symbolic alpha factors on crypto data, the other develops time-series models on US equities. They share no agents, memories, candidate spaces, or research state. Each part fixes its splits, feature and label definitions, and evaluator before any iteration begins. The agent emits only a constrained factor expression or a single config diff. The researchers call this asymmetric freedom: the agent explores freely inside its domain-specific language, but the evaluator sits outside the adaptive surface.

Part I is a six-agent pipeline orchestrated by an AI Manager. Agents never call one another directly; every handoff goes through the Manager, keeping runs auditable. A factor enters as a falsifiable proposal—hypothesis, mechanism, predicted direction, and refutation conditions—before being assembled from a registry of standard formulaic-alpha operators. Because each time-series operator reads only a trailing window and each cross-sectional operator reads only the current timestamp, causality is closed under composition. 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, 0.151 for AlphaGen, 0.137 for LSTM, 0.106 for LightGBM, and 0.075 for an Alpha158-style baseline. Individual factor ICs remain weak at 0.026–0.037, indicating the improvement comes from the harness, not a single expression.

Part II predicts each stock's forward return over the next 30 minutes on intraday US equities. Training runs on 2010–2019, 2020 is an embargo gap, and 2021–2025 is untouched test data. Selection uses only an inner-validation slice from the end of the training window. A hypothesis is one config diff—architecture, loss, sampler, or optimizer—and produces exactly one variant. The predictor is a hybrid featuring a multi-scale 1-D convolutional front-end, a configurable backbone (LSTM, Mamba, or attention), a cross-sectional stage, 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 from +0.0251 for ridge to +0.0843 for the hybrid, an absolute gain of +0.0230 over the best baseline, or 37.5% relative. The two parts' ICs use different conventions and should not be compared.

The per-stock score becomes a dollar-neutral threshold long/short book at a two-leg cost of 2 basis points. Sector-neutralizing raises the held-out Sharpe to +2.15. A causal volatility-targeting overlay lifts it to +2.50, and a fully causal walk-forward that chooses 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.

The two independent research loops share no agents, memory, or state. Freedom is asymmetric: the agent explores within its DSL, but the evaluator remains outside the adaptive surface. What improves is the research process itself.