Classical derivations of mean-variance preferences have all relied on the expected utility hypothesis. Numerous experimental studies have revealed that the expected utility model is systematically violated in practice. Such findings and the simplicity of the mean-variance framework have led researchers and practitioners to employ the mean- variance model without expected utility. However, the theoretical foundations of these models are scant.I provide behavioral foundations for a class of mean-variance preferences. My set of axioms characterizes an individual who assigns subjective probability to events and judges each portfolio solely on the basis of the mean and variance of its implied distribution over returns but does not necessarily rank the portfolios according to expected utility. I clarify the differences across specifications of my model. In addition, this model is robust to the consideration of a wide body of observed behaviors under uncertainty, which are inconsistent with the classical mean-variance model.