Earlier tests of asset pricing models are performed under the assumption of rational expectations and presume that the use of realized returns as a proxy for expected returns is acceptable. This paper turns the tables and asks what realized returns we would observe, given the pricing model. We develop an approach that opens new avenues in testing asset pricing models and explaining observed puzzles related to risk and return. This approach is used to test a long-standing puzzle: the price of market risk, lambda, as been found small, even negative and often insignificant. Our approach links the price of market risk to changes in conditional variance and its long-term persistence. We compare our price of market risk estimates with the ones estimated using the traditional approach. Empirically, we use both conventional measures of variance, such as (asymmetric) GARCH models against new measures based on forward-looking option market implied volatilities and the MIDAS (mixed data sampling) estimation. Tests are conducted using US stock market data. The results indicate that the lambda estimates from the new approach are consistently significant, positive, more stable, and higher than those estimated using the traditional approach. The new approach works even with return horizons shorter than one month.