Vlad Rodeski

An autonomous research agent was given nine published U.S.-equity anomalies and survivorship-free data (1998–2025, ~61,000 tickers including delisted names) and told to determine which effects are real. The study runs on two layers: a quant layer asking whether each documented effect reproduces on clean data, and an agent layer asking where the AI goes wrong and what checks catch the error before a wrong answer ships.

The one apparent “survivor” (out-of-sample Sharpe 1.95) was the agent’s own construction error, a mis-built sort, not a real anomaly. Checking the build against the source paper was the only thing that caught it. On clean data, faithfully built, none of the nine survived out-of-sample.

The agent’s raw output couldn’t be trusted alone — left unchecked it built a momentum strategy from a contrarian paper, applied wrong price screens, and guessed universes from paper titles. A two-pass fidelity review (code review + trade-log review) caught these.

Point-in-time data is decisive: sorting on retroactively split-adjusted prices manufactured a fake edge that compounded to ~500x, versus decaying toward zero on as-traded prices — about 3.5 Sharpe points of pure illusion from one careless price field.

Decay is the rule, not the exception. All three historically real effects (Bali, Frazzini, Chen) faded to insignificance or reversed over the 2023–25 out-of-sample window, and no long-only anomaly beat passive SPY.

Link to full article: https://polyquantai.com/research/equity-anomalies-replication