The proliferation of anomalies and the resulting `factor zoo' has challenged finance researchers to identify firm characteristics that are genuinely related to the cross-sectional variation in expected stock returns. We address this challenge using a Bayesian ensemble of trees approach, namely, Bayesian Additive Regression Trees (BART), which combines the advantages of machine learning with a Bayesian inference framework. Applying the methodology to U.S. stock returns and a large set of characteristics, we find a firm's market value is the sole consistently relevant characteristic. A BART framework that exploits the information content of a sparse set of characteristics identified in real time, offers substantial gains over competing models in both statistical and economic terms. We further confirm that the stochastic discount factor based on a sparse set of factors can successfully explain most of the variation in expected returns. A key reason for the sparsity of the factor set identified by BART is the methodology’s ability to incorporate non-linearities and high order interaction effects in a non-parametric manner. Our findings suggest that the vast majority of the documented anomalies are redundant, and in this sense our results offer a more optimistic view of the state of affairs in asset pricing.