Do T-Stat Hurdles Need to Be Raised?: Direct Estimates of False Discoveries in the Cross-Section of Stock Returns

Abstract

Recent studies argue that multiple testing casts doubt on cross-sectional stock return predictors. I show selective reporting casts doubt on certain multiple testing adjustments. t-stat hurdle adjustments require data on both null and non-null predictors, but predictors close to the null are unreported, leading to weak identification. In contrast, adjustments that target published findings focus on reported predictors and are more strongly identified. Accounting for identification problems in a dataset of 155 cross-sectional predictors, I find the data say little about whether t-hurdles should be raised, but at least 80% of published cross-sectional predictors are true (in-sample, before trading costs).