Much of the extant literature predicts market returns with “simple” models that use only a few parameters. Contrary to conventional wisdom, we theoretically prove that simple models severely understate return predictability compared to “complex” models in which the number of parameters exceeds the number of observations. We empirically document the virtue of complexity in U.S. equity market return prediction. Our findings establish the rationale for modeling expected returns through machine learning.

We propose a new asset pricing framework in which all securities' signals predict each individual return. While the literature focuses on securities' own‐signal predictability, assuming equal strength across securities, our framework includes cross‐predictability—leading to three main results. First, we derive the optimal strategy in closed form. It consists of eigenvectors of a “prediction matrix,” which we call “principal portfolios.” Second, we decompose the problem into alpha and beta, yielding optimal strategies with, respectively, zero and positive factor exposure. Third, we provide a new test of asset pricing models. Empirically, principal portfolios deliver significant out‐of‐sample alphas to standard factors in several data sets.