Does Academic Research Destroy Stock Return Predictability?
Source: McLean & Pontiff (2016) · The Journal of Finance · DOI: 10.1111/jofi.12365
TL;DR
Replicating 97 published anomalies, the authors show predictability decays ~26% after the original sample period (statistical bias / overfitting) and ~58% after publication — implying that publication itself, by alerting arbitrageurs, erodes roughly an additional one-third of an anomaly's return. This is the empirical cornerstone of out-of-sample skepticism and the reason platforms like ConvexPi grade strategies on hidden post-sample data.
The problem it addresses
Published cross-sectional anomalies are discovered and reported using in-sample data. Two forces should weaken them afterward: (1) statistical bias — with thousands of variables tested, the published ones are partly lucky, so returns shrink even on fresh data with no behavioral change; and (2) arbitrage — once published, sophisticated investors trade the signal and compete the premium away. McLean & Pontiff quantify and separate these two effects.
Main findings
Out-of-sample (post-sample, pre-publication) decay ≈ 26%. Comparing each anomaly's return after the original sample ends but before it is published isolates statistical bias / overfitting. Returns fall about a quarter — what you'd expect from selection among many tested signals.
Post-publication decay ≈ 58%. After publication, returns fall by more than half relative to in-sample.
The publication effect ≈ 32% (58% − 26%). The incremental decline attributable specifically to publication — i.e., to arbitrageurs learning about and trading the anomaly.
Cross-sectional pattern consistent with arbitrage: anomalies that are more arbitrageable decay more post-publication — those with higher in-sample returns, and those concentrated in larger, more liquid, lower-cost stocks.
Trading-activity evidence: after publication, the stocks an anomaly trades show higher volume, higher variance, and higher return correlation with other anomaly portfolios — fingerprints of arbitrage capital arriving.
Methodology
Hand-collect and replicate 97 anomalies from their original publications, reconstructing each signal and long-short portfolio.
Define three regimes per anomaly: in-sample (original study window), out-of-sample (after the sample ends, before publication), and post-publication (after the journal article appears).
Compare mean returns across regimes; use the post-sample/pre-publication window to net out pure statistical bias, attributing the additional post-publication drop to the publication/arbitrage channel.
Corroborate with post-publication changes in volume, volatility, and cross-anomaly correlation.
Implications for factor investing
Discount published backtests heavily. Expect roughly half of an anomaly's in-sample return to disappear in live trading; budget for it.
Out-of-sample evaluation is essential. The only credible test is performance on data not used (and not knowable) at strategy-design time — exactly the hidden post-sample evaluation ConvexPi enforces. In-sample Sharpe is not evidence.
Favor harder-to-arbitrage signals if you want persistence: smaller, costlier-to-trade segments decay less (though they are also harder to actually harvest).
Crowding is real and measurable. Rising volume/correlation among a factor's stocks is a warning that the premium is being competed away.
Key references
McLean, R. D. & Pontiff, J. (2016) — Does Academic Research Destroy Stock Return Predictability? — Journal of Finance — DOI: 10.1111/jofi.12365
Harvey, C., Liu, Y. & Zhu, H. (2016) — …and the Cross-Section of Expected Returns — Review of Financial Studies
Chen, A. & Zimmermann, T. (2022) — Open Source Cross-Sectional Asset Pricing — Critical Finance Review
Schwert, G. W. (2003) — Anomalies and Market Efficiency — Handbook of the Economics of Finance
Chordia, T., Subrahmanyam, A. & Tong, Q. (2014) — Have Capital Market Anomalies Attenuated in the Recent Era of High Liquidity and Trading Activity? — Journal of Accounting and Economics
