1. Publication: how much is lost the moment it is told

Level 10: Why it runs out

Every strategy anyone sells you comes with a date of birth and, almost always, with a date of death nobody mentions. This module is about measuring the second one. The question is not whether a return pattern was real when it was discovered, but what exactly happens to it on the day it stops being a secret.

An anomaly has three lives, not one

When a researcher documents a pattern, the work carries two different dates that almost nobody separates: the one on which the data used ends and the one on which the paper appears in print. Between the two there are usually three to six years, because reviewing and publishing is slow. That delay, which looks like a bureaucratic detail, lets you slice any pattern's history into three stretches:

• In sample. The period with which the author built and fine-tuned the result. It is the shop window: here the pattern always looks good, by construction.
• Out of sample, before publication. The years after the end of the study's data but before the paper circulates. The idea is no longer being fitted to that data, and nobody knows about it yet.
• After publication. The pattern is in the public domain. Anyone with capital can try to exploit it.

Comparing those three stretches turns an argument about opinions into a measurement.

The two numbers worth memorizing

McLean and Pontiff did exactly that with nearly a hundred patterns published in academic journals: they replicated them one by one and tracked their behavior afterwards. The two central results are these. Out of sample but still before publication, the return of those portfolios falls by around 26%. Once published, the fall goes to around 58% relative to the original result.

It is worth translating this into money, because in percentages of percentages the intuition gets lost. A pattern returning 6 points a year in the paper is left with about 4.4 points in the intermediate stretch and about 2.5 points once published. And those 2.5 points are before fees, before tax and before the cost of getting in and out. What started out looking like a money-making machine ends up being a narrow margin that a mediocre executor eats whole.

Two possible readings, and the stretch that separates them

Faced with that fall, two explanations are available, and they are very different from each other:

• The market learns. Publishing is teaching. Capital appears that buys the cheap and sells the dear following the recipe, prices adjust and the opportunity narrows. If this is what is happening, the pattern was real and has been arbitraged away.
• The result came inflated. The author tried many variants and presented the one that came out best; the journal, moreover, prefers striking results. Part of what was measured was not a regularity of the world, but luck that happened to fit that period. If this is what is happening, nothing has disappeared: there was never that much.

Here is the elegance of the design. The intermediate stretch distinguishes the two readings, because in those years nobody could have been arbitraging an idea that had not yet been told. Therefore that first 26% fall cannot be the market learning, and it has to be attributed to the original number coming inflated. The additional fall, the roughly 32 percentage points separating the 26% from the 58%, is compatible with informed money getting to work.

The answer, then, is not one of the two readings: it is both, in comparable proportions. Nearly half the erosion was already on the paper on the day of publication.

The traces learning leaves

If the market really does learn, it should leave visible traces beyond the return, and it does:

• The trading volume of the shares making up those portfolios rises after publication.
• The short interest in the leg the recipe says to sell grows appreciably.
• Portfolios built from already published patterns move more closely together, a sign that common capital is trading them.
• The erosion is not uniform: what is easy to exploit erodes far faster, that is, whatever lives in large, liquid, cheap-to-trade shares, and considerably less what requires touching small, illiquid or hard-to-borrow securities.

That last point is the hardest to explain by chance alone. If it were all statistical deflation, the erosion would have no reason to concentrate precisely where arbitrage is convenient.

What to do with this when you are shown a track record

Three practical consequences:

• A strategy backed by a famous paper is not safer for being so: it is older. Fame measures how many people already know.
• If the track record runs up to last year, ask what year the data used to design the rule ends. Without that date you do not know which part of what you are looking at is shop window.
• Insist on seeing the intermediate stretch and the post-publication stretch separately. A result that only shines in sample has not yet demonstrated anything.

Real case: the size pattern, published in 1981

The best documented case is the relationship between a company's size and its return, published in a 1981 paper with data going back to the 1930s. In the original sample the annual difference was several percentage points, an enormous result for the time. That same year the first commercial fund built explicitly to capture it was launched.

What came next illustrates the whole module: over the following two decades the size advantage was approximately nil, with long stretches even negative. Something similar happened to the Monday pattern, documented in 1980: it stopped showing up shortly after anyone could read about it. In neither case is there a clean answer about how much was the market learning and how much was a number that was too pretty. For anyone who bought the idea after reading about it, the result was the same.