Early in the firm's research we examined a deliberately naive proposition: that the documented anomalies of the academic finance literature, the calendar and cross-sectional regularities that fill decades of journals, could be harvested directly, in the unconditional form in which they are published. The turn-of-the-month effect, post-earnings-announcement drift, the January seasonal, the long roster of cross-sectional predictors, each arrives with a citation history and an in-sample t-statistic large enough to read as a standing invitation. The optimistic prior is that the literature is a menu of priced inefficiencies awaiting execution.

We evaluated it and chose not to pursue it. The reasoning is worth recording, because it recurs across the published cross-section.

Measured unconditionally, by whether the mean return remained reliably positive out of sample, the large majority had already reverted toward statistical insignificance by the time we tested them. Part of that gap is genuine decay; part of it was never there to begin with, the original significance inflated by the breadth of the search behind it and by costs the in-sample number ignored. The result associated with McLean and Pontiff frames the rest: predictor returns are lower out of sample, and lower again after publication. They read the initial out-of-sample decline as largely statistical, a product of overfitting and selection in the first sample, and the incremental post-publication decline as consistent with arbitrage by informed capital once an effect is named and disseminated.

Crowding is the leading explanation, not the whole one. A meaningful share of the published cross-section is marginal or negative once realistic frictions, borrow, and capacity are imposed, much of it living in hard-to-short, low-capacity names where gross and net diverge sharply. Much of the literature, then, is less a menu than a registry of premia that paid until enough capital knew they existed.

The more instructive finding lay one layer down. The effects were not uniformly dead; they were conditional. Some that appear flat unconditionally reassert themselves once returns are conditioned on state: a volatility regime, a phase of the cycle, a precondition in the recent path of prices. The unconditional anomaly had been arbitraged to noise; a narrower, regime-gated remnant persisted beneath it. But that remnant is a categorically different object: lower in breadth, less stable, more fragile, and contingent on a conditioning set one must specify correctly and re-estimate continually. Capturing it is a more exacting discipline than taking a published mean at face value.

So we retired the line in its stated form. The case against it was overdetermined: the unconditional premium had decayed, publication was associated with that decay in a manner consistent with arbitrage, and whatever signal remained lived in conditional structure rather than in the headline result, where identification, estimation error, borrow, and turnover levy their own tax.

The generalization is the part worth keeping. A documented edge and a durable edge are not the same quantity, and the most heavily documented effects are often the most heavily competed. Publication behaves as a decay function on expected return: the cleaner, more cited, and more readily replicated an effect, the more arbitrage capital it draws and the shorter its remaining half-life. Surviving predictability tends to be conditional rather than unconditional. An effect that works on average is often advertising that it is already crowded; an effect that pays only within a specifiable regime is the form that uncompeted return more commonly takes. What falls out is a working discipline: discount any result in proportion to how widely it has been read, and prefer the unglamorous, conditional, regime-aware specification of an idea to the clean unconditional headline that first drew attention to it.