Generalized Portfolio Sorts for Factor Validation

Abstract

Portfolio sorts are widely used in empirical asset pricing to identify firm characteristics that predict stock returns. However, such tests can conflate genuine characteristic-based predictability with persistent, firm-level heterogeneity. To address this limitation, we propose a Generalized Portfolio Sorts (GPS) model, which can exactly replicate results from all variants of conventional portfolio sorts, but can also be specified so that it separates a firm characteristic’s genuine predictive power from stable firm-level factors. We also derive a statistical test to detect whether return predictability arises from the sorting characteristic itself or from persistent, firm-level traits. Applied to a large set of proposed asset pricing predictors, we find that nearly half lose significance once persistent, firm-level heterogeneity is accounted for. The GPS-model thus strengthens factor validation, advances our understanding of the factor zoo, and provides a more robust foundation for empirical asset pricing tests.

Keywords: Portfolio sorts, cross-section of expected returns, tests of asset pricing models, random effects assumption

JEL Classification: C21, G14, D1