Lesson 5 of 8
What publication does to an edge
Ninety-seven published predictors, measured again after their papers appeared: returns were 58 percent lower after publication, 26 percent lower out of sample.
By Javier Sanz
Founder of ValueScreener. Built Ninety Nine, a securities services provider, and sold it to a tier-one global bank; ran operations and internationalization at Alpaca, the brokerage API.
Who writes thisEditorial processLinkedIn (opens in a new tab)X (opens in a new tab)
In one minute
- McLean and Pontiff took 97 published predictors of stock returns and measured them again after the paper that documented each one.
- Returns were 26 percent lower out of sample and 58 percent lower after publication.
- The difference between those two figures is their estimate of what publication itself does, as opposed to what data mining did.
A published edge has a problem no other kind of edge has: everybody can read it. That is testable, and it has been tested on a large sample of the same literature that the value premium belongs to.
What happens to a documented predictor, in three stages
- In the original sample100%
- Out of sample, before publication74%
- After publication42%
What they did
They collected 97 variables that published studies had shown to predict returns in the cross-section of stocks, then rebuilt each portfolio and measured it over two later periods: the years after the study's sample ended but before the paper appeared, and the years after publication.
Portfolio returns were 26 percent lower out of sample and 58 percent lower after publication. The split matters more than either number.
The two things a decline can be
Data mining
- The original result was partly luck found by searching.
- It shows up as a fall the moment you leave the original sample, before anybody has read the paper.
- Their 26 percent out-of-sample decline is the upper bound on this.
Learning and crowding
- Investors read the paper and trade the pattern, and the price adjusts.
- It shows up only after publication, on top of the out-of-sample fall.
- They estimate the extra decline at 32 percent, the difference between 58 and 26.
Their design separates them, which is the reason the paper is cited rather than the size of the decay.
The details that make it convincing
Three further findings in the same paper point the same way, and they are the parts a reader should carry.
Predictors with higher in-sample returns decline more after publication, which is what you would expect if the most attractive results attract the most money. Returns are higher for portfolios concentrated in stocks with high idiosyncratic risk and low liquidity, which are the stocks arbitrage capital finds hardest to work in. And after publication, a predictor's portfolio becomes more correlated with the portfolios of other published predictors, which is the signature of the same investors trading the same published list.
Their own conclusion is stated plainly: the findings suggest investors learn about mispricing from academic publications.
What follows for a person reading a screen
This is also the cleanest test between the two readings set out earlier in this course. A risk premium has no reason to shrink when it is written up. A mistake does.
One caution before the discount is applied to everything. Decay is measured across a population of predictors, and the population is dominated by recent, narrowly specified variables rather than by the handful of old, heavily replicated relations. A finding that was documented in the 1930s, rediscovered in the 1990s and re-tested by its own authors since is not in the same position as a characteristic published four years ago, even though both appear in the same average. The number tells you what to expect from the typical published edge, which is the correct default and not a verdict on any particular one.
Check yourself
How much lower were the portfolio returns after publication in McLean and Pontiff's sample?
- 26 percent
- 58 percent
- They were unchanged
Show the answer
Answer: B.
Returns were 26 percent lower out of sample and 58 percent lower post-publication, across 97 published predictors.What does the 26 percent out-of-sample decline mainly measure?
- The cost of trading
- An upper bound on how much of the original result was data mining
- The effect of publication itself
Show the answer
Answer: B.
The out-of-sample period comes before publication, so a decline there cannot be caused by readers. The authors treat it as an upper bound estimate of data mining effects.Which finding is the signature of investors trading the same published list?
- Higher returns in large, liquid stocks
- Post-publication increases in correlation between published-predictor portfolios
- Lower volatility after publication
Show the answer
Answer: B.
If the same money crowds into the same documented patterns, those portfolios start moving together, which is what the paper measures after publication.Sources
- McLean and Pontiff, "Does Academic Research Destroy Stock Return Predictability?", Journal of Finance, volume 71, number 1, February 2016: 97 variables, returns 26 percent lower out of sample and 58 percent lower post-publication, with a 32 percent estimate for publication-informed trading.
What this leaves out
- The figure's three bars restate the paper's percentages against a base of 100. The underlying levels are in the paper and are not reproduced here.
- The 97 variables are cross-sectional predictors of stock returns, which include but are not limited to value measures.
- A decline is not a disappearance. The paper measures how much smaller the returns became, not that they reached zero.
- The worked example is invented arithmetic on an invented 6 point premium, used to show the shape of the discount rather than to describe any rule.
- Nothing here is advice about any security or strategy.