Post-Publication Decay: What McLean and Pontiff Actually Measured

The trade gets smaller when it gets famous

A factor paper does not merely describe a market pattern. It can advertise the pattern to every reader with a Bloomberg terminal, a data subscription, and a mandate to find something that looks different from the benchmark. If the pattern is real, publication can make it less profitable. If it is partly statistical theater, publication can reveal that too.

McLean and Pontiff (2016) studied this process across 97 published predictors of the cross-section of stock returns. Their result is one of the cleanest empirical reasons to discount a backtest, since returns decline after the original sample, and they decline further after the anomaly becomes public knowledge.

There is one numerical trap worth fixing at the start. The paper’s headline estimates are a 26% decline in returns in the out-of-sample period and a 58% decline in the post-publication period relative to the original in-sample benchmark. The difference is 32 percentage points. That 32-point gap is the estimated effect associated with publication-informed trading, not a separate claim that the pre-publication out-of-sample decline itself is 32% (McLean and Pontiff 2016). The shorthand often gets swapped around in summaries. The underlying lesson survives the bookkeeping.

Three clocks, not one

McLean and Pontiff divide the life of an anomaly into periods that answer different questions.

The in-sample period is the period used by the original paper to identify the predictor. Its return is selected, at least partly, because it looked interesting. It is the least independent number in the story.

The out-of-sample period is later data that precedes publication, or otherwise is not used in the original discovery sample. This period asks whether the characteristic predicts returns before the academic result is widely disseminated. A decline here is consistent with selection bias, sample-specific luck, and ordinary estimation noise.

The post-publication period follows the publication date. It asks a harder economic question: what happens once investors can learn the signal from the paper? A decline beyond the pre-publication out-of-sample decline is consistent with arbitrage, crowding, and trading costs.

These periods should not be treated as interchangeable. “Out of sample” means not used in the original estimation. “Post publication” means after the information became available to the market. A test can be out of sample but not post-publication, and it can be post-publication while also being out of sample. The labels refer to different axes of information flow.

The empirical pattern

For each predictor, the authors construct a portfolio that is long stocks with high values of the characteristic and short stocks with low values, using conventions close to the original literature. They then compare average returns across the three periods. The sample begins in 1927 and the collection covers 97 predictors from published studies.

The average return falls in stages. Relative to the original in-sample estimate, the out-of-sample return is about 26% lower. Relative to the same benchmark, the post-publication return is about 58% lower. The publication-informed component is therefore about 32 percentage points. The exact decomposition is an estimate, not a physical meter installed inside the market, but it gives a useful scale: more than half of a reported anomaly return can disappear once the pattern is both tested on new data and known by investors.

The effect is not evenly distributed. Predictors with larger in-sample returns tend to experience larger post-publication declines. Predictors concentrated in illiquid stocks and stocks with high idiosyncratic risk also show stronger publication-related decay, which fits a limits-to-arbitrage story. A strategy that requires trading expensive, hard-to-short securities can look spectacular in a frictionless table and ordinary in a portfolio.

The study therefore speaks to two separate forms of disappointment:

• The original paper may have overstated the effect because it selected from noise.
• The remaining effect may have been competed away after publication.

A careful researcher needs to allow for both. Blaming every decline on market efficiency is as lazy as blaming every decline on p-hacking.

A simple return decomposition

Let $\mu_I$ be the mean in-sample return, $\mu_O$ the mean pre-publication out-of-sample return, and $\mu_P$ the mean post-publication return. Define the relative decay rates

and

The publication-associated increment is

If the in-sample mean is $0.50\%$ per month, the pre-publication out-of-sample mean is $0.37\%$, and the post-publication mean is $0.21\%$, then

and

The arithmetic is simple; the interpretation is not. The 32% is not “the amount caused by investors” in a perfectly identified causal sense. Publication dates are imperfect measures of information arrival, and market conditions move for many reasons. McLean and Pontiff use cross-sectional patterns and timing to make the publication explanation plausible, not to pretend that all other causes have been switched off.

Mechanism one: statistical selection

The first source of decay exists before anyone reads the paper. Researchers rarely choose an anomaly from a hat. They look at candidate definitions, sample periods, portfolio construction rules, and controls. The published specification is often the one that survived that search.

Suppose $X_i$ is the estimated return from specification $i$ and every specification has true mean zero. If the estimates have standard deviation $\sigma$, the expected maximum is approximately

for $N$ roughly independent trials. The published estimate contains the maximum-selection term. A new sample does not contain the same favorable noise realization, so the return falls even if the underlying relation has not changed.

This is why the first out-of-sample haircut is not evidence that markets immediately arbitraged the factor. It can be evidence that the original estimate was a winner’s estimate. Harvey, Liu, and Zhu provide the complementary cross-sectional calculation: once hundreds of variables have been published, a conventional t-statistic is not enough to establish a new discovery.

Statistical selection is also why a “replication” that simply reruns the same code on a slightly longer sample can be more informative than a new model. The original model is frozen. The researcher does not get another garden of forks to walk through.

Mechanism two: trading costs were hiding in the table

Many academic anomaly returns are reported before realistic implementation costs. That is not a moral failure; gross returns are useful for describing a relation. It becomes a problem when a gross, equal-weighted, high-turnover portfolio is sold as a net investment opportunity.

Let turnover be $\tau_t$, one-way cost per dollar traded be $c_t$, and gross portfolio return be $r_t^G$. A simple net return is

If publication attracts capital, trading can move prices and widen the effective cost of entering the long and short legs. Small-cap and illiquid anomalies pay a particularly high tax. A monthly gross spread of 0.35% can be economically meaningful before costs and disappointingly ordinary after a 0.20% round-trip cost.

Trading costs also alter the signal itself. If a factor’s portfolio requires buying the same names as many other factors, the marginal impact is not independent. The covariance of trading demands matters. In a crowded book, the cost of a rebalance is a function of what everyone else is trying to do at the same time.

Aronson’s evidence-based technical-analysis framework makes the same practical point in a different vocabulary: a rule is not validated by a favorable historical curve if the curve was built without a plausible execution model (Aronson 2006, Chapters 9-12). The market pays in spreads, impact, borrow, delay, and missed fills. It does not accept a citation as payment.

Mechanism three: publication-driven crowding

Publication changes the information set. Investors who did not know the characteristic can now test it. Funds can create products around it. Vendors can package it. Researchers can add it to a model and trade the names it selects. The market does not need perfect rationality for this to reduce returns; it only needs enough capital responding in the same direction.

A simple demand model makes the intuition clear. Let $q_t$ be the aggregate position in the anomaly and suppose the expected return is

where $\lambda>0$ measures how strongly crowding compresses the premium. Before publication, $q_t$ may be small and expected return close to $\alpha_0$. After publication, $q_t$ rises and the expected return falls.

The relation need not disappear. A risk premium may remain because investors dislike holding the underlying risk, because arbitrage is costly, or because capital is slow and segmented. “Crowded” does not mean “zero.” It means the original return estimate is an unsafe forecast.

The strongest crowding evidence appears where arbitrage is difficult. High idiosyncratic volatility, low liquidity, short-sale constraints, and high turnover all make a published anomaly expensive to exploit. The same features that created a large gross spread can make the post-publication implementation painful.

Mechanism four: genuine economic decay

Crowding can be temporary, but competition can also change the mechanism permanently. If a signal reflected slow information diffusion, and investors build systems that process that information faster, the original delay shrinks. If a premium compensated investors for bearing a risk, the premium may persist; if it was a behavioral mistake that became widely known, it may weaken.

A useful distinction is between discovery and deployment. Discovery puts the signal in a paper. Deployment changes prices when capital trades it. The effect can decay even if the paper is never read by a portfolio manager, because the paper enters teaching materials, data products, factor indexes, and other signals that transmit the idea indirectly.

Genuine alpha decay is difficult to estimate separately from costs and selection. McLean and Pontiff do not claim a neat causal split with laboratory precision. Their contribution is the timing pattern: the post-publication decline is larger than the earlier out-of-sample decline, and the cross-section of decay lines up with predictors that should be more vulnerable to arbitrage.

What the result says about market efficiency

The study is often summarized as “academic research destroys anomalies.” That is catchy and incomplete. It supports a version of market efficiency in which public information is incorporated gradually and imperfectly. If publication reduces a return, the market was not fully efficient before publication. If the return remains positive, the market is not instantly efficient after publication either.

The more useful question is not whether an anomaly is alive or dead. It is whether the mechanism behind it is still operating at a scale that pays after costs. A value signal can remain economically intuitive while its free-lunch version disappears. Momentum can persist in a portfolio while the exact lookback and skip rule from a thirty-year-old paper loses its edge.

For a working quant, the right prior is conservative: the original effect size is an upper bound, not a target. Forecast the haircut before you admire the backtest. If the strategy only works at the published magnitude, it probably does not work.

A decay-aware research workflow

Start by identifying the signal’s information age. A famous factor should carry a larger publication and crowding haircut than a genuinely new, difficult-to-implement signal. “New” is not the same as “unknown,” so use exposure data, factor correlations, vendor labels, and the number of products built around the idea where possible.

Then estimate the return through the implementation chain:

1. Gross return using point-in-time data.
2. Net return after spread, impact, borrow, financing, and turnover.
3. Capacity-adjusted return at the intended capital level.
4. Post-publication return in a frozen, independent sample.
5. Stress results across liquidity, market-cap, and shortability filters.

import numpy as np

in_sample = 0.0050
out_sample = 0.0037
post_publication = 0.0021
out_decay = 1 - out_sample / in_sample
post_decay = 1 - post_publication / in_sample
publication_gap = post_decay - out_decay
print(out_decay, post_decay, publication_gap)

Do not hide the haircut inside a single “expected Sharpe.” Show the gross-to-net bridge. Investors can argue about cost assumptions when they can see them; they cannot argue with a backtest that quietly omitted them.

The standard to carry forward

McLean and Pontiff do not say that published research is useless. They say that publication is an event in the life of a strategy. The in-sample number is what was once observed. The out-of-sample number is what survived a new sample before broad dissemination. The post-publication number is what remained after knowledge and capital arrived.

Use the 26%, 58%, and 32-point figures as a discipline for thinking, not as a universal haircut pasted into every model. A factor’s decay depends on liquidity, capacity, competition, signal half-life, and the quality of the original evidence. But do not start from zero decay. Start by asking why this one should be exempt.

The market does not punish you for reading a paper. It punishes you for treating the paper’s best historical estimate as a contract.

Turning decay into a forecast

A decay estimate is useful only if it changes the forecast. Start with the original gross return $\hat{\mu}_I$, then apply separate haircuts for selection, costs, and information diffusion rather than one mysterious discount factor. A simple planning equation is

The terms are not independent in practice. Crowding can raise costs, and an illiquid signal can have both a larger publication decline and a larger implementation cost. The equation is a forecast scaffold, not a causal identity. Its value is that it forces the researcher to state assumptions.

For example, suppose a paper reports 0.45% per month gross. A planning case assumes a 26% pre-publication decline, a further 32-point publication gap, and 0.10% monthly costs. The expected gross post-publication return is approximately

and net return is about 0.089% before financing and capacity. That is a very different investment proposition from the paper’s 0.45%. It may still be useful in a diversified portfolio. It should not be sized as if the original table were a live quote.

Run optimistic, base, and adverse cases. The optimistic case can use a smaller haircut for a low-turnover, capacity-rich signal. The adverse case should assume further decay when the signal is already crowded or when the paper’s sample ends near a favorable regime. If the strategy only works in the optimistic case, the research conclusion is fragile.

Monitor decay after deployment. Track rolling gross and net returns, exposure concentration, turnover, borrow, crowding proxies, and the gap between forecast and realized information coefficients. A falling return with stable exposures suggests competition or costs. A changing exposure distribution may indicate data drift. A sudden reversal around a publication or index launch is an information event worth investigating, not a reason to rewrite history.

There is also a portfolio-level lesson. If several strategies rely on related published factors, their publication haircuts are correlated. Applying a 30% discount independently to each factor can still overstate the combined portfolio because crowding arrives through common names and common trades. Stress the joint book, not only the individual sleeves.

The correct response to decay is not to abandon every public signal. It is to forecast the smaller, more expensive, more competitive version that investors can own. A strategy that survives that forecast has earned the right to be tested further.

• McLean, R. D. and Pontiff, J. (2016). “Does Academic Research Destroy Stock Return Predictability?” The Journal of Finance, 71(1), 5-32. doi:10.1111/jofi.12365.
• Harvey, C. R., Liu, Y., and Zhu, H. (2016). “…and the Cross-Section of Expected Returns.” The Review of Financial Studies, 29(1), 5-68.
• Aronson, D. (2006). Evidence-Based Technical Analysis. Wiley, Chapters 9-12.
• López de Prado, M. (2018). Advances in Financial Machine Learning. Wiley, Chapters 1, 7, and 11-15.