As AI agents become capable of autonomously conducting empirical research, a basic question arises: when can we trust the discoveries they produce? Autonomous systems can formulate hypotheses, implement and evaluate them, learn from intermediate results, and repeatedly redirect their search. The resulting hypotheses, their number, and their dependence structure are themselves outcomes of a random and highly adaptive research process, while multiplicity and information leakage can make spurious findings appear convincing. We develop a reusable statistical methodology for rigorous autonomous asset-pricing research that separates essentially unrestricted exploration from formal post-registration confirmation. A time-stamped boundary ledger records and freezes the realized research output and its confirmation designs; fresh post-boundary data are then reserved for statistical certification.
The key insight is that adaptive exploration need not compromise statistical validity once the realized search output is frozen and subsequent evidence is valid conditional on that output. We formalize this principle through conditional anytime e-closed testing for random, dependent hypothesis families. Our central result controls conditional expected loss at level $\alpha$, even when the reporting time is data dependent and the registered family and boundary-specified design choices are random. Different loss functions recover FDR, FWER, FDX, and other error criteria within a common framework; for indicator losses, the guarantees hold simultaneously across confirmation checkpoints.
We instantiate the exploration layer with SEADS, an autonomous system that proposes, implements, tests, and refines predictors of daily U.S. equity returns. Using two fresh years of post-registration data, we subject its discoveries to FDR- and FWER-controlled confirmation using the same evidence paths. Without being supplied known anomaly formulas or labels, SEADS reconstructs much of the established short-horizon structure of the equity cross section from primitive building blocks, including reversal, low-risk and volatility effects, turnover, and profitability. The results demonstrate how open-ended autonomous exploration can be coupled with rigorous fresh-data confirmation to produce statistically credible asset-pricing discoveries.
This is joint work with Xiaowei Ding (Nanjing), Jason Pan (Stanford), and Markus Pelger (Stanford).