This paper explores the theoretical and empirical implications of time-varying and unobservable beta. Investors infer factor loadings from the history of returns via the Kalman filter. Due to learning, the history of beta matters. Even though the conditional CAPM holds, standard OLS tests can reject the model if the evolution of investor's expectations is not properly modelled. We use our methodology to explain returns on the twenty-five size and book-to-market sorted portfolios. Our learning version of the conditional CAPM produces pricing errors that are significantly smaller than standard conditional or unconditional CAPM and the model is not rejected by the data.