1.10.4 Post-Publication Decay

Basic Prerequisites: 02 · The Factor Zoo and 03 · Crowding & Capacity.

---

1. Intuition & Practical Objective

The question every quant allocator eventually asks: if a factor’s premium is published in the Journal of Finance, does it still work afterwards?

The cleanest answer is McLean & Pontiff (2016), who assembled 82 characteristics from 68 studies, replicated 72 of them in-sample, and measured the premium in three windows: the original study’s sample, the period after the sample but before publication (out-of-sample, pre-publication), and the period after publication. Their logic is a natural experiment:

• If return predictability is entirely statistical bias (data snooping, sample selection, multiple testing), the premium should disappear out of sample - before anyone has even read the paper.
• If it is mispricing and arbitrage is costly, the premium should survive out-of-sample but decay after publication, when the research “draws attention to characteristics” and capital moves in.
• If it is a rational risk premium, it should survive both.

The findings are stark and quantitative:

• Out-of-sample decay ≈ 10%, and not statistically different from zero. You cannot reject that there is no statistical bias.
• Post-publication decay ≈ 35%, and statistically different from both 0% and 100% ($t\approx -4.9$ in their pooled specification). “An in-sample alpha of 5% is expected to decay to 3.25% post-publication.”
• Combining the two, “a lower bound on the publication effect of about 25%.”

And the mechanism shows up in the tape: “post-publication, stocks in characteristic portfolios experience higher volume, variance, and short interest, and higher correlations with portfolios that are based on published characteristics.” Crucially, “post-publication return declines are greater for characteristic portfolios that consist of stocks with low idiosyncratic risk” - i.e. the effect is strongest exactly where arbitrage is cheapest, which is the signature of mispricing being competed away rather than risk.

The one-sentence essence. “A published factor premium is an arbitrage with a half-life: roughly a third of it disappears after the paper comes out, the decay is largest where arbitrage is cheapest, and the out-of-sample decay before publication is too small to explain it - so the premium is mostly mispricing, and it is being harvested.”

Why it matters practically. It sets the discount rate you should apply to any factor’s backtest: multiply your in-sample alpha by roughly 0.65 before you believe it. It also implies the value of research is front-loaded: the earlier you trade a discovered factor (before it is famous), the more of the premium you capture - which is exactly why proprietary research and short-horizon signals are defended so fiercely.

---

2. Mathematical Ground Truth & Derivations

The decay estimator. Normalize each characteristic’s mean return in the post-sample periods by its in-sample mean:

$${\hat{d}}_{\text{OOS}}=1-\frac{{\bar{r}}_{\text{OOS}}}{{\bar{r}}_{\text{IS}}},{\hat{d}}_{\text{post}}=1-\frac{{\bar{r}}_{\text{post}}}{{\bar{r}}_{\text{IS}}}.$$

This normalization is the key design choice: it converts a heterogeneous collection of characteristics - some with 0.2%/month premia, some with 0.8% - into a common scale so they can be pooled. The pooled mean is the average survival fraction.

The pooled regression form. MP also estimate a specification with returns in levels,

$${\tilde{r}}_{i,t}=\alpha +{\beta }_{\text{OOS}}1[\text{post-sample}]+{\beta }_{\text{post}}1[\text{post-publication}]+\epsilon ,$$

where the coefficients, divided by the in-sample mean, are percentage decays. Their headline coefficients correspond to the ~10% / ~35% split, with the post-publication dummy significant at $t\approx -4.9$ and robust to clustering on anomaly, controls for time trends, SSRN posting dates, and - importantly - a linear decay in the post-publication months rather than a step.

The Fama–MacBeth / two-pass structure (why each characteristic is a portfolio). For characteristic $i$, they form a long-short portfolio each month, scaled by the in-sample means, and run a Fama–MacBeth cross-sectional regression of returns on the characteristics each month, then average the monthly slope coefficients. The three windows are compared by the same estimator. A characteristic enters only if its in-sample $t$-statistic exceeds 1.50 (their replication filter; note that 10 of 82 could not be replicated at all).

The breadth problem (why this result needed 82 characteristics). A single characteristic’s post-publication premium has enormous sampling error: the monthly portfolio volatility (≈3–5%) is an order of magnitude larger than the alpha (≈0.5%). Detecting a 35% decay in one characteristic would take literally centuries of data:

$$T^{\star }={\left(\frac{2\sigma }{\alpha \cdot d}\right)}^2={\left(\frac{2\times 0.03}{0.005\times 0.35}\right)}^2\approx 1176\text{ months}\approx 98\text{ years}.$$

It is only the breadth across 82 characteristics that makes the decay observable. This is Grinold’s law applied to research: the signal is weak, so the sample must be wide.

Reconciling with risk premia. A risk-based factor (say, a bond risk premium) is not expected to decay on publication - you cannot arbitrage away compensation for bearing risk. MP’s finding that the decay is concentrated in low-idiosyncratic-risk, liquid, dividend-paying stocks is precisely the pattern that distinguishes arbitrage-able mispricing from risk compensation: costs of arbitrage are lowest there, so capital arrives fastest.

---

3. Computational Implementation - the decay experiment

Stdlib only. This is McLean–Pontiff in miniature: 82 characteristics, an in-sample period, an out-of-sample pre-publication period, and a post-publication period, with the pooled survival fractions recovered.

The simulation recovers the published result: a 34.2% post-publication decay (MP: ~35%), an out-of-sample decay that is markedly smaller and less reliable, and the headline translation 5%/yr in-sample → 3.29%/yr post-publication (MP: 3.25%). The last line is the one that matters for research design: a single characteristic could not detect its own decay in under a century - the result exists only because MP pooled 82 of them.

This is the discriminating test. If publication decay were driven by a statistical artefact, it would be unrelated to arbitrage cost. Instead the decay is more than three times larger for the cheap-to-arbitrage group (48.7% vs 16.6%), and the difference is ~32 percentage points at $t=2.42$ - statistically distinguishable even in this small simulation. Expensive corners of the market keep their anomaly longer because capital arrives slower. That is the fingerprint of mispricing, and it is also the practitioner’s opportunity map: the durable factors are the ones that are annoying to trade.

---

4. Failure Modes & First-Principles Breakdowns

1. Backtests are biased upward by ~35%, and the bias is not uniform. Applying a flat 35% haircut is better than nothing, but the correct haircut is characteristic-specific: bigger for liquid, low-idio-risk factors, smaller for illiquid ones. A capacity-aware haircut is the honest default.
2. Replication failure is itself a finding. 10 of MP’s 82 characteristics could not be replicated in-sample - their own original papers’ results did not reproduce. Combined with the 35% decay, the expected realised alpha of a randomly-selected published factor is far below its abstract.
3. The decay is slow, not a cliff. MP find “a linear decay in predictability during the post-publication months,” not a step. This means (i) the effect is a diffusion of information and capital, not a switch, and (ii) the date of publication is an excellent instrument precisely because it is exogenous to the return series - which is why their estimate is credible where a simple pre/post comparison would not be.
4. Confusing decay with regime. A factor that underperforms after publication may be in a bad regime rather than decaying. MP control for time trends in characteristic returns and for “characteristic returns simply being lower during the later years in our sample” - a discipline any practitioner should copy before concluding a strategy is dead.
5. Decay and crowding are the same phenomenon on different clocks. Publication is a discrete, datable trigger of crowding; the crowding in 03 · Crowding & Capacity is the continuous version. A factor with a wide valuation spread and rising cross-factor correlation is in the late, crowded stage of the decay curve.
6. Do not over-read the 35%. MP explicitly can “reject the hypothesis that post-publication return-predictability does not change and we can also reject the hypothesis that return-predictability disappears entirely.” The premium shrinks by about a third; it does not vanish. Factor investing remains viable - just at a lower, capacity-dependent rate.

---

5. References

• McLean, R. David & Pontiff, Jeffrey: “Does Academic Research Destroy Stock Return Predictability?” (Journal of Finance, 2016)
• Cochrane, John H.: “Presidential Address: Discount Rates” (JF, 2011)
• Ilmanen, Antti: Expected Returns (2011) - “The persistent success of any asset class or trading strategy leads to a ‘virtuous’ cycle of growing popularity and further success, resulting in eventual overcrowding and subsequent disappointments.” Verified against the corpus book.
• Schwert, G. William: “Anomalies and Market Efficiency” (Handbook of the Economics of Finance, 2003)
• Jegadeesh, Narasimhan & Titman, Sheridan: “Profitability of Momentum Strategies: An Evaluation of Alternative Explanations” (JF, 2001)
• Bailey, Borwein, López de Prado & Zhu (2014) and Backtesting Hygiene & Deflated Sharpe

---

6. Connected Graph Bridges

• Back: 03 · Crowding & Capacity · 02 · The Factor Zoo · Index Hub
• Forward: 05 · Failure Modes & Practice · 06 · Advanced Extensions: Factor Timing
• Efficiency evidence: Event Studies (how market efficiency is actually tested) · Backtesting Hygiene & Deflated Sharpe
• The factor family: Momentum (the publication counter-example) · Fundamental Multi-Factor Models