We study the trade-off between signal selection and diversification in asset pricing when many return predictors are available. Using the data-mining framework of Yan and Zheng (2017), we form long–short portfolios from financial ratio signals and evaluate performance relative to the CAPM and the Fama–French six-factor model. Although null signals are prevalent, portfolio performance is largely insensitive to their inclusion. Portfolios restricted to the most statistically significant signals underperform more diversified strategies. Out-of-sample information ratios are highest at p -value thresholds between 5% and 10%, well above levels typically advocated for false-discovery-controlled inference. The results indicate that diversification is more effective than strict inference-oriented signal selection for portfolio construction.
Goto et al. (Fri,) studied this question.