AQuA: Preventing Look-Ahead Bias in Self-Improving Quantitative Trading Agents

Why look-ahead bias is the central risk for LLM-driven quant research

When an LLM agent proposes hypotheses, writes experiments, reads results, and decides the next step, a leakage that survives the first round can be encoded as "successful experience" and magnified in later rounds. The paper formalizes the question: can a quant research agent keep learning from experiments while being structurally unable to alter data splits, evaluation rules, or the final test window?

The AQuA answer: two independent closed loops + one closed sandbox

AQuA splits research into two fully isolated systems:

• Part I (factor discovery) runs on a 5-minute crypto universe across 5 exchanges, 2021-2025. A manager coordinates multi-agent pipelines that propose falsifiable economic hypotheses and assemble factor signals.
• Part II (model development) runs on a US-equity 30-minute return prediction task with a config-driven loop that trains comparable model variants.

The two systems share no agents, memory, candidate space, or research state. They share only one principle: validated experiments may influence the next research decision, while the contract that decides what counts as a valid experiment is fixed outside the loop. The paper emphasizes that "improvement" is improvement of the research process, not of model weights.

Each iteration follows the same closed loop:

1. Hypothesize from the current research state. 2. Build / train: turn the hypothesis into an executable factor expression or model config. 3. Evaluate on a validation set. 4. Validate: check look-ahead bias and single-regime dependence. 5. Select / combine validated evidence into the research state. 6. Update state for the next round.

The closed sandbox: making look-ahead un-expressible

The sandbox enforces four hard constraints:

1. Data splits are frozen before autonomous iteration begins. 2. Feature and label definitions are frozen before autonomous iteration begins. 3. The evaluator is frozen before autonomous iteration begins. 4. Agents can only operate through a restricted domain-specific language (DSL) whose expressions cannot read or modify the above.

The decisive trick is causality closed under composition: every time-series operator in the DSL reads only a trailing window, and every cross-sectional operator reads only the current timestamp. Any expression the agent composes is therefore causally correct by construction; it cannot express a normalizer or label that would peek into the future, not even by accident.

Selection leakage is handled separately:

• During search, the agent receives only fixed validation-set score signals.
• The true test window is evaluated exactly once, after configuration freeze, and never returned to the loop.

The paper calls this arrangement "asymmetric degrees of freedom": the agent can explore freely inside the constrained DSL, while the evaluator sits outside the adaptive surface.

Headline numbers and their strict boundaries

Part I (crypto factor system)

• Composite factor signal IC ≈ 0.190 on the 5-minute crypto universe.
• Mixed data from 5 crypto exchanges, 2021-2025.

Part II (US equity model system)

• Single-name IC = +0.0843 on US 30-minute return prediction.
• Strongest baseline (GRU) = +0.0613.
• Absolute lift +0.0230, relative lift 37.5%.

Part II long/short threshold strategy

• Out-of-sample Sharpe = +2.50 with 2 bps round-trip costs.
• Positive yearly returns 2021-2025.
• Strict walk-forward causal evaluation: Sharpe still around +2.0.

The paper explicitly distinguishes 2.50 (validation / mixed evaluation) from 2.0 (stricter causal rolling evaluation) and does not present 2.50 as the final public number. That honesty strengthens the credibility of the results.

"The spy in the code" — AQuA's own near-miss

The most instructive section is a reverse case study. An earlier, freer version of AQuA gave the agent direct access to write feature code, with another LLM reviewing the code before execution. The team still saw data leakage.

The specific failure:

• The agent created an intraday-volume feature.
• The intent was to measure cumulative volume since the open and compare across stocks.
• The bug was in the normalization: cumulative-to-now volume divided by full-session total volume.
• At 10:00 AM, the numerator (open to 10:00 AM) is available, but the denominator (open to close) is not yet knowable.
• The model had quietly read a sliver of the future.

This is the classic shape of look-ahead bias: backtests look implausibly good because they use information that would not exist at trade time. The AQuA team did not hide this episode; they published it as direct evidence for why a DSL plus closed sandbox is necessary.

Why this is a watershed for AI quant research

The paper reframes what counts as "AI quant research." The field has moved through three stages:

1. Copilot era: researchers write logic, AI fills in syntax. 2. Alpha mining era (Alpha-GPT, QuantaAlpha): AI searches a restricted symbolic space, but human researchers make the selection judgment. 3. Autonomous research era (AQuA): AI proposes hypotheses, trains models, reads results, and decides the next step; humans only design the sandbox boundary.

AQuA is what makes stage 3 concrete rather than aspirational. It answers three real questions with empirical numbers:

• Can the agent avoid peeking at the future? DSL plus causal closure makes peeking un-expressible.
• Can selection leakage be controlled? Validation / test split plus "test evaluated once, never returned."
• Does recursive self-improvement self-contaminate? Five years of positive yearly returns plus walk-forward Sharpe still near 2.0.

Complementarity with the late-August Agentic Trading stack

AQuA sits at the research methodology layer, complementing infrastructure-layer pieces like Mint-Agent (financial-native agentic base model, 9B / 27B checkpoints), Binance Agent OS (MCP-compatible agentic wallet plus skill hubs), and Waton AlphaSchema (auditable semantic trading plans). The infrastructure layer lets agents run with compliant identity and auditable outputs; AQuA lets those agents produce strategies that can actually be trusted. Together they mark a shift in 2026 H2 from "tooling AI" toward "research-methodology AI."

| Layer | Agentic Trading stack | AQuA | |---|---|---| | Infrastructure | Mint-Agent / Binance / Waton | — | | Research methodology | — | Dual-system closed sandbox | | Error control | Wallet / sub-account / auditability | DSL causal closure + test isolation | | Headline evidence | 100M+ commercial orders | Sharpe 2.50 / IC 0.190 / 5-year positive returns |

Implications for AI for Science

The AQuA pattern is portable beyond finance:

• The multi-agent factor system generalizes to molecule generation, materials combinatorial search, and chip placement.
• The config-driven model loop generalizes to protein conformation search and molecular dynamics parameterization.
• The closed sandbox plus DSL generalizes to any domain where certain invariants must be made un-violable, such as physical conservation laws or irreversible thermodynamics.
• The causal closure property (operators cannot read outside their causal scope) becomes a reusable constraint language for trustworthy self-improving agents in science.

The core transferable principle: in every AI for Science domain, the first step toward letting AI run its own research is making sure AI cannot cheat.

Open questions the paper leaves on the table

1. Cross-system knowledge transfer. Part I and Part II are correctly isolated, but can insights from the factor system indirectly inform the model system? The current design forbids it. 2. Multi-market causal-closure DSLs. AQuA validates only on crypto and US equities. Writing equivalent DSLs for A-shares, Hong Kong stocks, futures, and FX is the main localization hurdle. 3. AI-designed sandboxes. Humans still design the DSL set, evaluator, and data split. Whether letting an AI pick from a fixed DSL family further reduces error remains an open engineering question.

What to do if you want to reproduce or extend AQuA

1. Reproducing AQuA requires preserving the full set of quantization choices, sampling settings, system prompts, tool schemas, and complete state-transition logs. Missing any of these means the numbers are no longer "AQuA's numbers." 2. Comparing different LLMs under AQuA requires three ablations: swap only the agent LM with the harness fixed; reset the research state with the LM fixed; change the evaluator contract with both LM and state fixed. 3. Porting to a non-finance domain should start by deliberately constructing a "spy" feature that would leak the future in your domain, then proving that the DSL plus causal closure actually blocks it. That is the cheapest engineering experiment for verifying closedness.

Sources

1. arXiv 2608.12841 v2, "AQuA: Recursively Self-Improving Quantitative Trading Research Agents," submitted 2026-08-13, updated 2026-08. 2. Princeton + Ant Group + Stanford collaboration (Jiacheng Guo, Suozhi Huang, Yunlong Gao, Zihao Li, Jason Ge, Xu Kuang, Mengdi Wang). 3. Pith Science T2 audit commentary. 4. Generative AI Publication case study on the "spy in the code" failure. 5. Emergent Mind full-paper analysis. 6. Mobile Tonghuashun Finance Chinese-language deep report, 2026-08-19.