Post-publication decay
Do stock-market factors still work after publication?
Mostly not the way the papers claimed. The finding — from McLean & Pontiff’s 2016 study and reproduced independently here — is that a published return anomaly earns far less after the world knows about it. On a replication of 18 of the most-cited factors, the average premium falls by roughly three-quarters from its in-sample strength, and only about five or six clear an honest multiple-testing bar.
Two reasons a factor fades
- Arbitrage. A documented mispricing is an invitation. Funds pile into the trade, prices adjust, and the gap the paper measured shrinks. This is the healthy explanation — markets getting more efficient.
- Data-mining. Hundreds of predictors have been tested. If you run enough regressions, some cross the significance line by pure chance. Those were never real, and they revert toward zero the moment you leave the original sample window.
What survived, what didn’t
The clearest survivors on this replication are momentum, industry momentum, betting-against-beta, and quality-minus-junk — signals with long records and a plausible risk or behavioural story. The casualties include most value definitions, accruals, the small-cap size effect, and several low-volatility measures, which either decayed sharply or never cleared the higher significance bar that the sheer number of tested factors demands.
Explore all 18
The zoo scatter plots every factor by publication year and replication t-statistic; toggle it to count only returns earned after publication, and watch most of them collapse toward the axis.
Open the zoo →Common questions
How much does a typical factor decay after publication?
McLean & Pontiff (2016) estimated returns fall about 26% out-of-sample and 58% post-publication on average. A replication of 18 of the most-cited factors here lands near a 75% average drop from in-sample strength to what they earned after their paper came out.
Why do they decay?
Two forces. Arbitrage: once a mispricing is documented, money floods in and trades it away. And data-mining: with hundreds of predictors tested, some were never real to begin with — they looked significant by chance and revert to zero once you're out of the original sample.
Which ones held up?
Momentum and industry momentum, betting-against-beta, and quality-minus-junk are the clearest survivors on this replication. Most value definitions, accruals, the size effect, and several low-volatility measures did not clear a multiple-testing correction.
Educational replication, not advice. Reproducing a premium in-sample is not evidence it will persist; figures move as the data updates.