Stats › Long run returns › Factor replication
The factor zoo and replication
Academic finance has spent sixty years hunting for things that predict which stocks go up. It has found plenty. Hundreds, in fact, and that abundance is itself the problem. When enough researchers test enough ideas against the same set of returns, some of them clear the bar by luck alone, and the ones that clear it are the ones that get published.
This page is about what happens when somebody goes back and checks. One team counted the published factors. Another retested hundreds of them with more careful portfolio construction and found most of them didn't survive. A third looked at what happens to the survivors once their paper is out in the world. The three results stack into an uncomfortable picture, and I've put them on one page because they usually get cited separately and they're far more useful together.
This isn't an argument that nothing works. Several factors do hold up and the page names them. It's an argument about how high the bar has to be before a backtest counts as evidence, and about which questions to ask of anyone showing you one. Every figure here comes from the working papers linked at the foot, and where the published versions differ from them I've said so on the page rather than smoothing it over.
The academic literature has published 316 factors claimed to predict stock returns across 313 papers. When Hou, Xue and Zhang retested 447 published anomalies using New York Stock Exchange breakpoints and value weighted returns, 286 of them, or 64%, were insignificant at the conventional threshold. At the stricter threshold that Harvey, Liu and Zhu recommend for new factors, t above 3, the failure rate rises to 85%. The authors of the factor count put it plainly: most claimed research findings in financial economics are likely false. And of the ones that are real, published returns fall 58% after publication.
Members only
The rest of this page is for members
Below this point there are 8 sections, 1 chart and 4 named sources, roughly 1600 words of it. Every figure carries the source it came from and the date the data is from.
You can keep browsing every statistic in the library for free. The intro and the summary are always open.