Artificial intelligence can now propose trading ideas, write experiments, compare models, and revise its own research agenda. That sounds like the beginning of an autonomous hedge fund. It also sounds like a machine built to automate backtest overfitting at industrial speed.
A new paper from researchers affiliated with Princeton, Stanford, and Ant Group confronts that danger. AQuA contains two independent research systems. One discovers crypto factors. The other develops models for US equities. Each remembers what worked, what failed, and why, then uses that evidence to shape the next experiment.
The results are striking. The crypto system reaches a combined information coefficient of roughly 0.190. The equity model produces a per-stock IC of 0.0843 and a held-out Sharpe as high as 2.50 after costs. Yet the deeper idea is architectural: if an AI researcher cannot reliably avoid leakage, remove its ability to create leakage.
