Looking Under the Hood of Data-Mining

This paper re-evaluates academic research on 92 cross-sectional stock return predictors. Researchers studying return predictability must make decisions about portfolio construction; for example, whether to rebalance annually or monthly. In sample, the returns of predictor portfolios constructed with the precise research decisions made in the original papers are significantly larger than those constructed with a random combination of decisions made in the literature. Out of sample, half of this difference disappears. The effects exist only for predictors published in top-ranked journals. The results suggest that statistical biases from research decisions explain a fifth of the return predictability in the literature