Persistent Anomalies and Nonstandard Errors

This article presents a framework for rigorous inference that accounts for the many methodological choices involved in testing asset pricing anomalies. We demonstrate that running multiple paths on the same original dataset inherently results in high correlation across outcomes, which significantly alters inference. In contrast, path-specific resampling greatly reduces outcome correlations and tightens the confidence interval of the estimated average return. Jointly accounting for across-path and within-path variability allows the variance of the average return to be decomposed into a standard error, a nonstandard error, and a correlation term. In our empirical analysis, we identify 29 persistent anomalies for which the 95% confidence intervals of their average returns exclude zero. Our tests also indicate that, for most anomalies, nonstandard errors dwarf standard errors and, in turn, are the primary determinants of the width of confidence intervals for multi-path average effects.