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 goodn ess-of-fit statistics that reject the restrictions too often. An itera ted GMM approach has superior finite sample properties. The coefficien t estimates are approximately unbiased in simpler models, but their as ymptotic standard errors are understated. Simple adjustments for the s tandard errors are partially successful in correcting the bias. In mor e complex models the coefficients and their standard errors can be hig hly unreliable. The power of the tests to reject a single-premium mode l is higher against a two-premium, fixed-beta alternative than against a conditional Capital Asset Pricing Model with time-varying betas.