We reexamine the robustness of the inference from the least-squares estimator under hetero-skedasticity and autocorrelation of unknown form in a generic multifactor asset pricing model. It is shown that the asymptotic covariance matrix of the least-squares estimator of betas depends only on the long-run cokurtosis of factors and error terms, whereas that of alphas depends not only on the long-run cokurtosis but also on the long-run coskewness of factors and error terms. We numerically evaluate the celebrated Fama-French three-factor model using the U.S. data and find considerable changes in sizes of asymptotic variance estimates of the least-squares estimator of alphas and betas due to nonnormality and serial dependence.

multifactor model, heteroskedasticity and autocorrelation of unknown form, asymptotic variance, alpha, beta.