Publishing an Anomaly Is How You Kill It

Every quant has a private worry. You find a beautiful signal, you check it carefully, it works. Then you notice that a paper described it four years ago. Is it still there? Or did the paper kill it?

R. David McLean and Jeffrey Pontiff turned that worry into a research project, and produced one of the most quietly important papers in modern finance. They tracked 97 published stock return predictors and asked what happened to each of them after the world found out.

The problem: three reasons a backtest looks better than reality

Suppose an academic publishes a signal that earned 8% a year in their sample. You test it today and it earns 4%. Why the gap? There are three completely different possible explanations, and they have completely different implications:

1. Statistical luck. The signal was never real. The researcher tested many things, and this one got lucky in their particular sample. Under this story, the signal never worked, before or after publication.
2. Sample-specific truth. The signal was real but only in that period, for reasons that have since changed.
3. Arbitrage. The signal was real, the paper told everyone about it, traders piled in, and their trading pushed prices to where the profit no longer exists.

These matter enormously. Under (1), the academic literature is largely garbage. Under (3), the literature is correct and the market is genuinely learning from it. Same observed decay, opposite conclusions.

The key idea via analogy: cut the timeline into three pieces

McLean and Pontiff's method is clean, and it is the reason the paper is convincing. For each of the 97 predictors, they carved the history into three periods:

1. The original sample period, the years the paper actually studied.
2. The post-sample, pre-publication period. The gap between the end of the researcher's data and the day the paper appeared. This is a genuinely out-of-sample period, but crucially nobody knew about the signal yet. No trading could have taken place on it.
3. The post-publication period. The signal is now public. Everyone can trade it.

That middle window is the stroke of genius, because it separates the explanations that would otherwise be tangled together.

• If the decay were purely statistical luck, the returns should collapse in period 2. Out-of-sample data does not care whether a paper has been published; a fake signal fails as soon as you leave the original sample.
• If the decay were purely arbitrage, returns should hold up in period 2 (nobody is trading it yet) and only fall in period 3 (now they are).

The results:

• Returns were 26% lower out-of-sample (period 2 versus period 1). So some of the original result was indeed luck and overfitting. Not none, but only about a quarter.
• Returns were 58% lower post-publication (period 3 versus period 1).
• The difference between those two, roughly 32%, is the decay attributable specifically to publication.

That 32% is the number the paper is famous for. It is a direct measurement of how much profit the act of telling everybody destroys.

They back this up with corroborating evidence: after publication, the stocks involved in an anomaly see higher trading volume, higher short interest, and their returns become more correlated with other published anomalies. That is exactly the fingerprint of arbitrage capital arriving to trade the signal.

Why it mattered

• It vindicated the academic literature, and indicted it, in one stroke. About a quarter of the average published anomaly's return was overfitting. But the rest was real, and it was real enough that trading it away worked. Both critics and defenders of the field got something.
• It showed markets actually learn. This is a striking result for the efficient market hypothesis. Prices are not born efficient, they become more efficient, and one of the mechanisms is people reading journals. Mispricing gets corrected because someone published a description of it.
• It gave alpha decay a number. Before this, "signals decay after publication" was a folk belief among practitioners. Now it is a measured effect with a magnitude, which is a completely different kind of knowledge.
• It changed how research is used. If you are building a strategy from a published paper, you should expect roughly half the paper's return, at best. That single planning assumption is worth more than most backtests.

The honest limitations

• Halved is not eliminated. A 58% decline still leaves 42%. Many anomalies remain profitable after publication, which is itself a puzzle: if arbitrage works, why does it not work all the way? The answer is presumably that arbitrage has limits, costs, and capacity, but the paper does not fully resolve it.
• The 97 predictors were the ones that got published. The vast graveyard of tested-and-failed signals is invisible, so the sample is inherently selected on success. The true decay across all ideas anyone ever tried is unknowable.
• Publication is not the only thing that happens. Technology improved, costs fell, and markets got faster over the same decades. Disentangling "the paper was published" from "everything got more competitive" is hard, and the paper's clean identification helps but does not entirely settle it.
• It cannot tell you which anomalies survive. It gives averages. Your specific signal may be dead, or may be one of the robust ones. The paper offers a prior, not a diagnosis.

The one-line takeaway

McLean and Pontiff tracked 97 published return predictors and found that about a quarter of their profit was never real, and another third was arbitraged away once the paper came out, which means academic research genuinely does make markets more efficient by destroying the very patterns it describes.