AQuA Makes Quantitative Trading Agents Self-Improving While Keeping Backtests Honest
New AQuA system from Princeton, Ant Group, and Stanford keeps backtest results reliable.
Quantitative researchers rarely celebrate an unusually beautiful backtest curve. They first ask whether future data leaked into the feature, whether the test set has been repeatedly viewed, or whether the returns come from a specific market regime. This professional habit reflects a basic truth: producing a good result is easy; proving it deserves belief is hard. When an agent starts proposing hypotheses, running experiments, and deciding next steps on its own, the problem becomes more serious. An unnoticed data leak can be saved as a success, and the agent may drift further along a false conclusion.
AQuA consists of two independent parts. Part I researches symbolic factors; Part II researches trainable models. They use different agents, memories, and candidate spaces but follow the same research principle: experiments that pass validation enter the next decision round, while data paths and evaluation rules remain outside the loop.
In experiments, Part I achieved a combined validation-set Spearman IC of about 0.190 on five-minute cryptocurrency data. Part II reached a per-stock IC of +0.0843 on 30-minute-ahead U.S. equity returns. A dollar-neutral long-short strategy built from Part II's signals delivered an out-of-sample Sharpe as high as +2.50 after including 2 basis points of round-trip transaction costs, with positive Sharpe in every year from 2021 to 2025.
Recursive self-improvement in AQuA does not change underlying model weights; it changes the research process. Before an experiment, the system forms a testable hypothesis. For factor research, it must specify the economic mechanism, the predicted direction, and the falsification condition. For model research, it translates the idea into a configuration change. After the experiment, the system saves not only the IC or Sharpe but also why the candidate was proposed, under which market conditions it held, how it compared with simple baselines, and what evidence contradicted the initial judgment. Later research resumes from these records.
The design was shaped by an early failure documented in the paper's appendix. Initially, an agent could directly write feature code, and another agent was responsible for checking look-ahead bias. One candidate feature, 'intraday volume participation rate', used total daily volume as the denominator. At midday, that denominator included volume from the rest of the day that had not yet occurred. Both the authoring and reviewing agents believed the feature used only historical information, but the code accessed future data. The leak produced an unusually high held-out IC, which did not survive re-splitting the evaluation window. This incident convinced the team to abandon the idea of adding another model-based review and instead restrict agents to pre-registered causal operators and a fixed set of available fields. The full-day total volume was not among them.
AQuA enforces two types of isolation. The first concerns the generation process: features, labels, and data transformations cannot access information after the prediction time. The data split, features, labels, and evaluator are fixed before the autonomous loop starts. Part I can only combine registered time-series and cross-sectional operators; Part II can only modify model structure, loss function, sampler, or optimizer settings. Models can choose research directions but cannot rewrite the data loader or evaluation logic. The second isolation concerns the selection process. During the search phase, agents only see a pre-defined validation score, used for candidate ranking, early stopping, and checkpoint selection. The final test window is run only after the configuration is determined, and its results are not returned to the agent for further tuning. In Part II, 2010–2019 is used for training, 2020 is an isolation band, and 2021–2025 is the final test window.
Part I is orchestrated by an AI Manager with six specialized agents, including a Data Steward, a Visual Analyst, an Idea Miner, a Factor Evaluator, a Backtest Engineer, and a Research Librarian. In one example, the system investigated whether a price rebound after a rapid decline in open interest tends to fail if there is no active buying or basis repair. The agent first looked for forced deleveraging events, compared subsequent changes in price, volume, basis, and taker flow, then turned the observation into a falsifiable mechanism. Candidate factors had to beat simple price, open-interest, and order-flow baselines, and also had to show that the effect concentrated in the hypothesized event window. Over 20 rounds of autonomous research, the combined validation Spearman IC rose to about 0.190 from single-factor ICs of roughly 0.026–0.037. The improvement came not from a sudden secret formula but from the gradual accumulation of weak signals with different economic mechanisms.
Part II tackles a different task: predicting each U.S. stock's return over the next 30 minutes. Every round submits a single config diff, making clear whether the change affected model structure, loss function, or training strategy. The final model uses a hybrid temporal architecture: multi-scale one-dimensional convolutions extract local price-volume patterns, a temporal backbone handles longer-range dynamics, and a cross-sectional module mixes information across stocks at the same time. Under the same evaluator, the strongest baseline GRU achieved a per-stock IC of +0.0613, while the searched hybrid model reached +0.0843. After industry neutralization, a dollar-neutral long-short portfolio had an out-of-sample Sharpe of +2.15 after transaction costs; adding a volatility control using only historical data raised it to +2.50. In walk-forward evaluation, where all parameters are determined by data before each test block, the Sharpe remained about +2.0.
The two parts cannot be directly compared because they use different evaluation metrics. But they jointly validate the same research methodology: evidence is allowed into memory, while evaluation stays independent. The approach can be transferred to other fields where an agent repeatedly calls an evaluator, such as hyperparameter tuning, materials discovery, drug screening, or model optimization. It also changes the human role: researchers no longer need to judge every inference step, but instead focus on defining data, acceptable experiment actions, comparison rules, and which results qualify for the next round.