We examine the empirical implications of learning under ambiguity for the cross-section of stock returns. We introduce a theoretically-motivated ambiguity measure and find that ambiguity is priced in the cross-section of average stock returns. Ambiguity is not subsumed by state variables known to predict stock returns, nor by value, size, and momentum factors. In R-squared comparative tests, a model that takes ambiguity into account performs better than empirical implementations of the Bayesian learning model, the intertemporal CAPM, and the four-factor model of Fama and French (1993) and Carhart (1997). (JEL G12) Conditional asset pricing models based on rational expectations perform poorly empirically (Lewellen and Nagel 2006; Kan, Robotti, and Shanken 2013). While the rational expectations hypothesis assumes investors know the probability law governing asset retu rns, authors such as Keynes (1921), Knight (1921), Shackle (1949), and Roy (1952) have emphasized that investors form expectations based on vague information that cannot be quantified precisely. Related evidence from experimental studies (Ellsberg 1961) has confirmed that individuals are averse not only to uncertainty regarding the outcome of events with known probabilities (risk), but also uncertainty regarding the outcome of events with unknown probabilities (Knightian We are especially grateful to the editor, Wayne Ferson, and an anonymous referee fortheir extensive comments that substantially improved the article. We also appreciate the valuable comments of Damian Damianov, Andrei Nikoforov, Jaime Casassus, and participants at the 2011 regional meetings of the Econometric Society in Santiago, Chile, the 2010 FMA meetings in New York City, and seminar at UTPA, Texas. Send correspondence