Hundreds of factors have been published.
Harvey, Liu and Zhu catalogued over three hundred in their 2016 study of the cross-section of expected returns and argued that the conventional significance threshold has become meaningless in a literature of that size (check the exact count against the paper; checked 2026-09-07). Hou, Xue and Zhang's 2020 replication of 452 anomalies reported that a large majority failed to clear the conventional threshold once microcaps were handled with New York Stock Exchange breakpoints and returns were value-weighted (check the exact proportion against the paper; checked 2026-09-07).
Why does testing 300 factors need a correction?
The arithmetic behind that is not subtle and you should be able to do it in your head. Test 300 strategies that have no edge at all, at a 5% significance level, and you expect 15 of them to look significant. The probability that at least one of the 300 clears the bar is 0.99999979. A t-statistic of 2.4 has a two-sided p-value of 0.0164, so the probability that at least one of 300 pure-noise trials produces a t of 2.4 or better is 0.9930.
Publishing that t-statistic without saying how many things were tried is not a small omission.
Two corrections are worth carrying. The Bonferroni critical value divides the significance level by the number of tests: with 300 tests at 5%, the two-sided critical value rises from 1.96 to 3.7648. It needs no assumption at all about how the tests relate to one another, which is what makes it usable on a literature nobody coordinated, and it is conservative for that same reason. Factor tests overlap heavily, so three hundred of them buy far fewer than three hundred independent chances, and Bonferroni charges you for all three hundred anyway. The sharper procedures, and the one that does assume independence, belong to the inference material.
Advertisement
What hurdle and haircuts should a new factor face?
The second correction is more intuitive. The expected maximum of N independent standard normals is approximately sqrt(2 ln N), which for N = 300 is 3.3775 and for N = 20 is 2.4477. The single best of three hundred noise strategies has an expected t-statistic above 3, which is why Harvey, Liu and Zhu proposed a hurdle around 3.0 rather than 2.0, and why a factor at t = 2.4 is a curiosity rather than a finding.
The same arithmetic converts into money. Over a twenty-year backtest a t-statistic of 2.4 corresponds to an annual Sharpe ratio of 0.5367, while a hurdle of 3.0 demands 0.6708. Insisting on the higher hurdle is a 20.0% haircut to the Sharpe ratio you are allowed to believe, applied before any transaction cost.
Post-publication decay is the empirical confirmation. McLean and Pontiff's 2016 study found that predictor returns fall materially out of sample and fall further after publication, with the post-publication decline larger than the out-of-sample one (check the exact percentages against the paper; checked 2026-09-07). Apply the two haircuts to a 5.0% in-sample premium and it becomes 3.7% out of sample and 2.1% after publication. Whether the cause is arbitrage by readers or overfitting by authors, the number you should plan around is the last one.
How should you read a new factor claim?
So read a new factor claim the way you would read a sell-side target price. Ask what universe it was tested on and whether microcaps drive the result. Ask whether returns are equal- or value-weighted, since equal weighting quietly concentrates the bet in the smallest names. Ask what the sample period is and what happened after it. Ask how many specifications were tried before this one was reported, and treat "we only ran this one" as an unverifiable claim. Ask whether the factor survives once the factors you already own are controlled for, which is the orthogonalisation check the signal-construction work already teaches.
And ask who is telling you, because a factor with an exchange-traded product attached to it has a sponsor.