Publication Bias and the Cross-Section of Stock Returns

Abstract

We develop an estimator for publication bias and apply it to 156 hedge portfolios based on published cross-sectional return predictors. Publication bias adjusted returns are only 12% smaller than in-sample returns. The small bias comes from the dispersion of returns across predictors, which is too large to be accounted for by data-mined noise. Among predictors that can survive journal review, a low t-stat hurdle of 1.8 controls for multiple testing using statistics recommended by Harvey, Liu, and Zhu (2015). The estimated bias is too small to account for the deterioration in returns after publication, suggesting an important role for mispricing. * We thank Rebecca Wasyk for excellent scientific programing and Chris Hollrah for excellent research assistance. This paper originated from a conversation with Frank Schorfheide. We also thank

Common Risk Factors in Cryptocurrency

“…We consider a comprehensive list of the established factors in the cross-section of stock returns, compiled by Feng et al (2017) and Chen and Zimmermann (2018). Among these, we select all the factors that can be directly constructed using only the information on price, volume, and market capitalization.…”

The Limits of p-Hacking: A Thought Experiment

“…Thus, by the central limit theorem, the sample mean return is approximately normal and the t-stat is approximately standard normal. Chen and Zimmermann (2018) show that this approximation holds very well for a 312 month sample of equal-weighted long-short quintile portfolios sorted on B/M. Equation 1also assumes that performance is uncorrelated across predictors, consistent with the near-zero average pairwise correlation between monthly long-short returns Though t-stats are on average zero, published t-stats are large due to authors' and journals' preferences for large t-stats.…”

“…This table summarizes the data and provides moments used in the estimation. CZ replications is the 156 replications of equal-weighted long-short quintile portfolios in Chen and Zimmermann (2018). HLZ estimated model simulates the model from Harvey, Liu, and Zhu (2016) I calculate the probability that a random t-stat is published (Equation (6)).…”

“…I estimate a model of pure p-hacking (Equations (1)-(2), circles) on large datasets of published t-stats (bars) by method of moments (Section 2.2). The top panel uses 156 replicated long-short portfolios from Chen and Zimmermann (2018). The bottom panel uses moments from Harvey, Liu, and Zhu (2016).…”

“…To rigorously pursue the p-hacking theory, I write down a statistical model in which factors have no explanatory power, but published t-stats are large because the probability of publishing a t-stat t i follows an increasing function p(t i ). I estimate p(t i ) by fitting the model to the distribution of published t-stats in Harvey, Liu, and Zhu (2016) and Chen and Zimmermann (2018). The p-hacking story is powerful: The model fits either dataset very well.…”