Conditioning Information, Out-of-Sample Validation, and the Cross-Section of Stock Returns

Abstract

Empirical research on conditional asset pricing has been built on several standard return-predictive variables. However, recent studies have raised serious doubts on these variables that typically serve as the instruments to capture the relevant conditioning information. In the stochastic discount factor framework, we propose and implement a new approach to assess the value of the standard instruments. We compare the out-of-sample performances of conditional models that are built on different subsets of several widely-used instruments. We find that some combinations of these instruments, after adjusting for the effect of the horse-race over all the subsets, can significantly improve the out-of-sample performance for pricing the cross-section of stock returns. In contrast, some other subsets give rise to conditional models that drastically underperform the unconditional model. The results affirm the value of the conditioning instruments for cross-sectional asset pricing and highlight the importance of instrument selection.