184 ideas judged, two alive, and both share one property: someone is compelled to trade on a visible schedule. Start here, because it is the claim everything else tests.

The cleanest compelled flow there is: leveraged funds must rebalance at every close, in a size arithmetic dictates. Two tests, both rejected, and the rejection is the interesting part.

Why it was rejected. The flow is real and the edge is real, and both live inside the cost of trading them, at a frequency daily prices cannot reach.

Where the obligation is heading next. When an idea reaches millions of accounts in ninety seconds, the crowd itself becomes the counterparty that has to trade.

Three value-at-risk models, 4,922 out-of-sample days, all three fail their coverage tests while the regulatory traffic light still reads green. A passing test is not a working model.

One level deeper: the dataset itself may be a model output wearing a measurement's clothes. A fitted value cannot carry information its inputs never had.

So how do you test honestly? Lock the decision rule before the first run. A 24.7-million-parameter model then loses to a four-parameter regression on 21 tickers out of 21.

The end of the road: duration measures cleanly on Treasuries and degrades on credit, and convexity turns out not to be measurable from daily prices at all. Two review rounds reversed this paper's conclusion twice before it published.

A published anomaly inverts its sign in a survivor-only sample, because the firms that carried the information are the ones that failed and got deleted.

The same deletion, one level up. A high yield index whose weakest borrowers migrate to private credit stops measuring the asset class and starts measuring what stayed.

And what it costs you in practice: five tests of the trade that seeded modern quant, where the surviving profit turns out to be a liquidity fee earned in the names hardest to trade.

Crowded positions do move together more than they should, and they do underperform in bad months. Neither result clears the bar to become a signal.

The same question asked of short interest. Four declared attempts, four failures, and one stress-state result that survives as a warning rather than a trade.

So what does crowding actually cost a portfolio? What imperfect correlation buys, and what it stops buying at exactly the wrong moment.

Build the security and simulate it: a three-tranche securitization of AI-infrastructure debt, 200,000 paths, with 2008's four named defects mapped onto today's complex.

Then check the collateral against the physical world. Announced gigawatts are positions in a queue, and the queue does not care how the deal was financed.

The whole view, priced. Fourteen pre-registered forecasts on what stays scarce when intelligence gets cheap, with four prints named in advance that would kill the thesis.