False Discovery Control for Factors

If a researcher tests 200 candidate factors at the usual 5% significance level, roughly 10 of them will look statistically significant purely by chance, even if none of them actually predict returns. This is the multiple-testing problem applied to asset pricing, and Giglio, Liao and Xiu built tools to handle it properly at the scale the factor zoo actually operates at, thousands of candidate signals, many of them correlated with each other rather than independent tests.

Their approach adapts false-discovery-rate control (originally from genomics, where the same problem of testing thousands of genes at once arises) to asset pricing's specific quirks: returns are noisy, factors are correlated, and the number of candidates keeps growing as researchers publish more. The output is an adjusted significance bar, much stricter than the ordinary 5% cutoff, that keeps the expected share of false "discoveries" among reported significant factors under control.

Worked example. A researcher runs significance tests on 500 candidate signals at the standard 5% level and finds 30 "significant" ones. Applying false-discovery-rate control at a 5% target, the adjusted threshold shrinks the surviving list to just 6 signals, the other 24 are judged likely artifacts of testing so many candidates at once, not genuine return predictors.