Linear Approximations and Tests of Conditional Pricing Models

We construct a simple reduced-form example of a conditional pricing model with modest intrinsic nonlinearity. The theoretical magnitude of the pricing errors (alphas) induced by the application of standard linear conditioning are derived as a direct consequence of an omitted variables bias. When the model is calibrated to either characteristics sorted or industry portfolios, we find that the alphas generated by approximation-induced specification error are economically large. A Monte Carlo analysis shows that finite-sample alphas are even larger. It also shows that the power to detect omitted nonlinear factors through tests based on estimated risk premiums can sometimes be quite low, even when the effect of misspecification on alphas is large.