FINITE-SAMPLE PROPERTIES OF THE GENERALIZED-METHOD OF MOMENTS IN TESTS OF CONDITIONAL ASSET PRICING-MODELS

We develop evidence on the finite sample properties of the Generalized Method of Moments (GMM) in an asset pricing context. The models imply nonlinear, cross-equation restrictions on predictive regressions for security returns. We find that a two-stage GMM approach produces goodness-of-fit statistics that reject the restrictions too often. An iterated GMM approach has superior finite sample properties. The coefficient estimates are approximately unbiased in simpler models, but their asymptotic standard errors are understated. Simple adjustments for the standard errors are partially successful in correcting the bias. In more complex models the coefficients and their standard errors can be highly unreliable. The power of the tests to reject a single-premium model is higher against a two-premium, fixed-beta alternative than against a conditional Capital Asset Pricing Model with time-varying betas.