We develop a new methodology for estimating time-varying factor loadings and conditional alphas based on nonparametric techniques. We test whether long-run alphas, or averages of conditional alphas over the sample, are equal to zero and derive test statistics for the constancy of factor loadings. The tests can be performed for a single asset or jointly across portfolios. The traditional Gibbons, Ross and Shanken (1989) test arises as a special case when there is no time variation in the factor loadings. As applications of the methodology, we estimate conditional CAPM and Fama and French (1993) models on book-to-market and momentum decile portfolios. We reject the null that long-run alphas are equal to zero even though there is substantial variation in the conditional factor loadings of these portfolios.