What the Original 61-Strategy Audit Taught Me About AI and Markets.
The field report that began with 61 strategies remains a dated research snapshot. The public corpus has since grown to 133 repositories; its lesson about where AI adds value in markets, and where it does not, still holds.
The title is deliberately blunt. Not every sleeve failed, but most original theses changed materially once they met real data, realistic costs, unstable correlations, and the market's tail. Several ideas survived only after the exciting story was removed. What remained was often smaller and less glamorous: a risk overlay, a sizing rule, a diversified ingredient, or a warning not to trade.
The most counter-intuitive lesson was that better prediction was rarely the durable advantage. Geography repeatedly collapsed into US beta plus currency risk. Sophisticated wrappers often became ordinary exposure with an additional fee. Low correlation disappeared during stress. A higher backtested Sharpe sometimes came from concentration in the period's winner, not from a more resilient portfolio. The useful finding was frequently a constraint.
This is where AI helps. It makes the cost of asking another question dramatically lower. It can inspect more hypotheses, translate an idea into reproducible code, find inconsistencies across documents, challenge parameter choices, and accelerate the comparison between a claim and its evidence. Used properly, AI expands the surface area of falsification. It helps a disciplined researcher discover more reasons to say no before capital has to do it.
AI does not decide which problem is worth solving. It does not know whether an apparent relationship is already priced, whether the available proxy represents the real economic exposure, or whether a result can survive liquidity, taxes, crowding, and a regime not present in the sample. Most importantly, it cannot own the consequences of a decision. Those are framing, judgment, and governance problems.
What surprised me most was how often the research answered the wrong question correctly. An optimizer could identify the best historical weights without showing that those weights were stable. A model could classify stress after it arrived without providing a tradable lead. A simulation could improve average return while relocating risk into one unrepeated event. In each case, computational success concealed decision failure. The hard work was not producing the answer. It was defining what an answer had to earn before anyone could act on it. That bar cannot be delegated to the model.
The corpus changed my view of intelligence in markets. Intelligence is not the volume of ideas a system can produce. It is the quality of the questions surrounding a decision: What would disprove this? Which risk is being relocated rather than removed? What happens in the worst five days? Who can stop the process? What evidence earns promotion from research to validation, and from validation to live?
The master lesson: AI is most valuable as an instrument for structured doubt. It can make research faster, broader, and more internally consistent. It becomes dangerous when speed is mistaken for understanding, confidence for causality, or generated output for accountable judgment.
That is why the graveyard stays public. The rejected sleeves show the framing process, not just the winners left after selection. They make the surviving system more credible because the record exposes what was tested, what broke, and why. Markets do not reward the appearance of intelligence for long. They reward decisions that remain coherent after the story fails.
