I develop a stochastic growth model with production where there is a hidden state governing productivity growth regimes, and the hidden state evolves according to a Markov chain. Economic agents learn about the hidden state and display ambiguity aversion in the spirit of Klibanoff et al. (2005). I calibrate the model to the post-war US data. The main findings are (1) the model can generate a high and volatile equity premium while a low and smooth risk-free rate, (2) agents' fluctuating beliefs induce countercylical variation in equity premium and in the expected volatility of returns, and moreover volatility clusterng and persistence; and (3) Bayesian learning itself is unable to generate a significant and positive risk premium once time variation in investment opportunities is accounted for; in most cases, Bayesian learning lowers the unconditional mean of equity premium.